Training Example: AI//COST – Review the Data, Give Your Score & Compare to the Real AI Evaluation

[H1]
The
independent price
desk for
every
AI model
on the market
Every rate is read from the vendor's own price page and stamped with the date it was read. API tokens for every major lab, plus the subscriptions, AI agents and calculators that turn a rate into a bill.
35 Providers live 343 Models tracked 50 Comparisons shipped 61 Capability Q&A shipped § 02 / INDEX
[H2] The full desk.
Sitemap → Section Coverage Pages Updated Status Pricing Hub → Per-model pricing with calculators and historical charts 343 pages Oct 30 LIVE Comparisons → Two-way head-to-heads with live pricing, benchmarks, and verdicts 50 pages Oct 30 LIVE Capability Q&A → "Can X do Y?" with verified test outputs and screenshots 61 pages Oct 3 LIVE Calculators → LLM API cost, token counting, subscription vs API, video generation, text-to-speech, embeddings, fine-tuning, agent cost and image generation — on live rates 9 live tools Oct 31 NEW AI Agents → Every route is priced from the vendor's page, our own test runs, and where each tool fails 12 tools / 36 pages Oct 30 NEW Glossary → AI terms, each defined in a sentence and tied to what it costs on live rates 38 terms Oct 21 NEW The Journal → Editorial: news cycle, deep analysis, opinion, research digests 20 posts Oct 28 LIVE
[H2] Flagship rates
6 LABS · CURRENT FRONTIER TIER · RATES VERIFIED 2026-10-31 All models → Anthropic Claude Opus 5.5 $4/M Input $20/M Output $2.58/M Blended 1M ctx
· cache $0.20/M SEE DETAILS →OpenAI GPT-6 Astra $10/M Input $50/M Output $6.82/M Blended 1.05M ctx
· cache $1.00/M SEE DETAILS →Google DeepMind Gemini 4 Argon $2/M Input $10/M Output $1.21/M Blended 1M ctx
· cache $0.10/M SEE DETAILS →SpaceXSI Grok 4.7 $2/M Input $6/M Output $1.19/M Blended 500K ctx
· cache $0.50/M SEE DETAILS →Moonshot (Kimi) Kimi K3.1 Pro $2/M Input $10/M Output $1.28/M Blended 1M ctx
· cache $0.20/M SEE DETAILS →CHEAPESTZhipu (Z.ai / GLM) GLM-5.3 $1.40/M Input $4.40/M Output $0.78/M Blended 1M ctx
· cache $0.26/M SEE DETAILS →
Blended rate assumes a 92% input / 8% output, 82% cached workload — the same assumption behind every
effective-rate figure on this site. Raw input and output rates are read off each
vendor's own pricing page. Where a vendor publishes no cached-input rate, no cache
discount is applied, so that model's blended figure is a ceiling rather than a
like-for-like number.
[H2] Featured comparisons
SOURCE-VERIFIED · LIVE PRICING & BENCHMARKS All comparisons → § COMPARE / 012026-10 Grok FreevsSuperGrok Grok Free ($0) gives limited Chat, Imagine and Grok Build. SuperGrok closed to new customers on October 12,… Choose Grok Free for casual chat. § COMPARE / 022026-10 ChatGPT FreevsChatGPT Paid ChatGPT Free gives unlimited text chat on GPT-5.6 Luna but meters every tool — uploads, images, voice, deep… Choose Free for casual chat. § COMPARE / 032026-10 ChatGPT GovsChatGPT Plus ChatGPT Go ($8/mo) is Free with higher tool limits — more messages with tools, uploads, images and memory — on… Choose Go for cheap headroom. § COMPARE / 042026-10 ChatGPT PlusvsChatGPT Pro ChatGPT Plus ($20/mo) and Pro ($100, $200 or $500) share the apps and the GPT-6 models in Work and Codex. Pro… Choose Plus unless you hit its limits weekly or need Pro reasoning. § COMPARE / 052026-09 GrokvsChatGPT ChatGPT leads the neutral scores — Intelligence Index 52.7 vs 46.5, Terminal-Bench 4.0 59.1% vs 25.8% — and… Live pricing, benchmarks, and a verdict by use case. § COMPARE / 062026-09 GrokvsClaude Both sides moved in September: Grok 4.7 shipped on Grok 4.6’s card, Claude Opus 5.5 shipped under Opus 5’s.… Live pricing, benchmarks, and a verdict by use case. § 03 / JOURNAL
[H2] Latest from the desk.
All entries →TOOLS · HOW-TOSEP 27 · 17 MIN
[H3] input_tokens means three different things. Your cost function knows one of them.
openaianthropicgemini↗TOOLS · HOW-TO · FEATUREDSEP 19 · 17 MIN
[H3] A 429 is free. The ninth retry is not.
rate-limitsapi-errorsprompt-caching↗ANALYSIS · FEATUREDAUG 31 · 24 MIN
[H3] The list price is not the price your agent pays
prompt-cachingcost-optimizationagents↗ANALYSIS · FEATUREDAUG 27 · 22 MIN
[H3] Running many agents at once: the tax is linear, the rebate is quadratic
claude-codeagentssubagents↗TOOLS · HOW-TO · FEATUREDAUG 23 · 19 MIN
[H3] The 20x in Claude Code caching is real — it's a penalty, not a discount
claude-codeprompt-cachinganthropic↗TOOLS · HOW-TO · FEATUREDAUG 21 · 16 MIN
[H3] The Batch API discount isn't 50% -- and it isn't always a discount
batch-apicost-optimizationapi-pricing↗
[H2] The tape doesn't lie.
Every price move, model release, and policy change in the AI market — logged within hours, not weeks. → Subscribe to the wire OCT 26 GPT-6.1 Terra launched — $1 / $5 per M tokens. RELEASE OCT 24 Kimi K3.1 Pro launched — $2 / $10 per M tokens. RELEASE OCT 19 MiniMax M3.1 Flash launched — $0.15 / $0.6 per M tokens. RELEASE OCT 19 Mercury 3 launched — $0.2 / $0.7 per M tokens. RELEASE OCT 16 Glimpse launched — $0.6 / $3 per M tokens. RELEASE OCT 12 Muse Wonder launched — $4 / $20 per M tokens. RELEASE § 04 / DESK NOTES
[H2] How we actually work.
Full methodology → 01
[H3] Vendor-canonical, not press releases
Pricing data is verified manually against each vendor's own canonical pricing page — 23 frontier providers, no scraping, no third-party aggregators. Every number on this site has a verification date stamped next to it. 02
[H3] Models documented, tools shipping
Per-model pricing pages cover 23 frontier providers, and the first calculator — LLM API cost — is live on the same verified data. Side-by-side comparisons are next. We don't claim coverage we don't have. 03
[H3] No silent edits. Ever.
When a number is wrong, the page is fixed and the correction goes into a public log: date, page, what was wrong and what is right now. Y Independent operator. Founder verifies every claim before publish. Real LinkedIn linked from author profile.
Founder bio → Featured onTiny Startupstinystartups.com
[IMG: Featured on Product-Tower]
[IMG: Featured on IndieHunt]
[IMG: Launched on Launch Streak]
[IMG: Featured on LaunchFill]
[IMG: Launched on AI Website Directory]
CSS Crème — Web Design Inspiration & Design Systems
[IMG: Featured on ToolListed]
[IMG: Verified on DANG!]
[IMG: Featured on CodeHype]
[IMG: Listed on The Runway - LaunchFree.io]
[IMG: AI//COST — Launching soon on Awesome Indie]
AI Pulse - AI Tools Directory
[IMG: Featured on AI Tools for Marketers]
[IMG: Featured on Tolodora]
[IMG: Featured on LaunchNests]
[IMG: Featured on MakerHunt]
[IMG: Powered by Startup Fast]
[IMG: Featured on LaunchKiwi]
[IMG: Featured on Noonlaunch]
[IMG: Featured on Launch Llama Tools]
[IMG: Featured on LaunchBoosts]
[IMG: Featured on MakerHunt]
[IMG: AICost on Better Launch]
[IMG: Featured on Buildrship]
Free AI Tools
[IMG: AICost (AI//COST): LLM API Price Desk on ToolVerified]
[IMG: Listed on List Maro]
[IMG: AICost on PeerPush]
[IMG: AICost on Verified Tools — AI & SaaS tools directory]
[IMG: Featured on AITrustList]
[IMG: AICost - Featured on Founder.best]
[IMG: Listed on Kick Product]
[IMG: Featured on Shipstry]
[IMG: Listed on ProofStories]
[IMG: Powered by SubmitMySaas]
[IMG: Featured on ProductInspire]
[IMG: On PaveLaunch]
Listed on AgentConcierge
[IMG: Featured on Wall of Tools]
FridayHuntfridayhunt.com
[IMG: Featured on Go Publicly]
[IMG: Featured on AIGCLIST]
[IMG: Featured on Altern]
[IMG: Listed on Enlisted]
[IMG: Submit AI Tools]
[IMG: Featured on Next Launch]
[IMG: Featured on Wonderlaunch]
[IMG: Featured on ShipX]
[IMG: Featured on Wayfindio]
[IMG: Featured on MPAT]
Listed on PostLaunchKit
[IMG: Featured on IndieLaunch]
[IMG: Featured on VibeCodingList]
[IMG: AICost — Verified by KittyLaunch]
[IMG: Featured on Share Your Startup]
[IMG: Featured on DiscoverLaunches]
[IMG: Featured on ToolsAIApp]
[IMG: Featured]
[IMG: Featured on Upperstory]
[IMG: AICost on Siteefy]
[IMG: Featured on FastLaunch]
[IMG: Featured on Product Launchify]
[IMG: Featured on StartupWiki]
[IMG: Featured on StackLedge]
[IMG: Launched on MarketingDB]
[IMG: Featured on LemonLaunch]
[IMG: AICost badge]
[IMG: Launching today on ProductPerch]
[IMG: Featured on LaunchZone]
[IMG: Listed on domainrating.lol]
[IMG: Featured on Saasgrave Launches]
[IMG: Featured on Smartees]
[IMG: Featured on Spot Startups - A Free Product Hunt Alternative]
[IMG: Listed on AgentLocker — AICost]
[IMG: LaunchClash]
[IMG: Pro Launch Featured Badge (light theme)]
[IMG: AICost on OnlyFoundrs]
[IMG: Featured on IndexOf.AI]
[IMG: Featured on Traffiklist]
[IMG: Find AICost on ProgrammerNeeds]
[IMG: Ranked on OpenRank.bid]
[IMG: Featured on Roozna]
Listed on Tool Index
[IMG: Featured on PeplioHUB]
Tiny Startupstinystartups.comCSS Crème — Web Design Inspiration & Design SystemsAI Pulse - AI Tools DirectoryFree AI Tools Listed on AgentConcierge FridayHuntfridayhunt.comListed on PostLaunchKitELFind us on EarlyLaunchListed on Tool Index

Last verified 2026-07-20 PUBLISHING PRINCIPLES METHODOLOGY v1.3 REVIEWED QUARTERLY PUBLIC + AUDITABLE
[H1]
How we work
The operating manual for everything we publish. How pricing data gets into the snapshot, where benchmark scores come from, how rankings are scored, how head-to-heads are built. Written so anyone can audit our claims — and challenge them when we're wrong.
Methodology version v1.3 Reviews per year 4 Cool-off pass 48h Verification Manual Corrections log live → § 01
[H2] Editorial principles.
#principles
Six rules govern everything we publish. They are locked into the methodology — changing them requires a quarterly review with a written rationale, not an editorial decision. The rules below are the surface; the rest of this document is how they're operationalized.
01
[H3] Verify before asserting
Anything that is not a well-known constant must be verified at the time of publication. Prices change, APIs deprecate, benchmark scores update. We don't quote from memory or training data. Each claim has a timestamp and a source. 02
[H3] Score before ranking
Rubrics are locked before testing. We define criteria + weights publicly, then run tests. Scores for one candidate are not visible while another is being scored — each pass is blind to the others. Rankings are a function of scores, not preferences. 03
[H3] Honesty over diplomacy
If a tool is bad at something, we say so. Every brand profile has a "Don't use it for" section. If we'd recommend a competitor for a use case, we link to the competitor — even when our affiliate relationship is with the subject. 04
[H3] Sponsored, always labeled
If a sponsor paid for it, it says so. Sponsored placements carry a dashed warm border and an "ⓘ SPONSORED" tag, so they are never mistaken for editorial. 05
[H3] Corrections are public
When we get something wrong, we log it openly. The corrections page lists every correction with date, page, what was wrong, and what's now correct. We don't silently edit. We don't gaslight readers about what we used to say. 06
[H3] Methodology is versioned
This document is v1.3. Major changes ship as version bumps, each logged with a date and a summary at the foot of this page — so you can see exactly what shifted in our approach and when. No silent rewrites of how we work. ★ THE BACKSTOP If you find a violation of any rule above, email [email protected] with the URL and the issue. We respond within 5 business days and either correct, push back with reasoning, or escalate to a public correction. Every correction we issue is logged at /corrections/ — that page is the canonical source of truth for our track record. § 02
[H2] How pricing data is collected.
#pricing-data
Pricing for 343 LLM models across 35 providers is verified manually against each vendor's own canonical pricing page. No scrapers. No automated billing-API pulls. No third-party aggregators. The trade-off is honest: slower updates than a scraping pipeline would deliver, but every number has a human who checked the source URL on the date stamped next to the price.
01
[H4] Canonical source identified
For each model, the vendor's own canonical pricing page is the source of truth (e.g., anthropic.com/pricing, openai.com/api/pricing). Press releases, blog posts, and third-party aggregators are never used as source — they're often stale or out of date the moment a vendor ships a change. 02
[H4] Manual verification
Each price is fetched from the canonical URL via WebFetch, read against the published rate card, and the input rate, output rate, cache rate, and any tier thresholds are recorded by hand into the model's JSON record. One human, one source URL, one verification date — no parser layer between the vendor page and the published number. 03
[H4] Diff against snapshot
The new reading is compared against the previous published snapshot. Any change triggers a re-read of the full pricing page (not just the row that moved) to make sure the rest of the rate card didn't change at the same time. Mismatches are reconciled before the new price is published. 04
[H4] Publish + log
Approved changes propagate to all dependent pages on the next build. The price-history table records the change with timestamp, old value, new value, and the source URL that was verified. Every price on the site has a verifiable provenance trail. Target latency from a vendor announcement to a published correction is under 24 hours; trigger is a vendor-blog signal or a reader-reported discrepancy. Source typeExampleRefreshVerification Vendor pricing pageManual WebFetch read of anthropic.com/pricing, openai.com/api/pricing, etc.On change✓ Manual Vendor announcementOfficial blog/social signal triggers immediate re-read of the canonical page~24h target✓ Manual Currency exchangeEvery price is shown in USD. Where a vendor publishes only in its home currency — mostly Chinese labs quoting yuan — we convert at one dated ECB reference rate for the whole catalogue, so rows stay comparable: 6.7209 CNY/USD as of Aug 28, 2026. A vendor that publishes its own USD list is read directly and never converted, because the two are separately set and do not track any exchange rate. Per-country display currency is still Phase 2.Quarterly✓ Manual Country tax/VATStatic config sourced from official government pages (planned — Phase 2, country pages not yet shipped)—~ Planned Third-party aggregatorsListed as not used as source; only used as a sanity cross-check—✗ Not source § 03
[H2] Where you actually buy.
#access-points
The same model is usually sold through several doors: the vendor's own console, a cloud marketplace, a consumer subscription app, sometimes open weights you can host yourself. Each provider page lists those routes because the door you walk through changes what you pay — and the rate we publish belongs to exactly one of them.
The published per-token rate is always the first-party rate, read from the vendor's own pricing page as described above. Where a route bills differently — a marketplace markup, per-seat licensing, a region premium — the route's card says so rather than silently reusing the direct number.
RouteWhat it isRate on the cardIn the dataset Direct console / APIThe vendor's own platform — console.anthropic.com, platform.openai.com. The route our published rates describe.First-party✓ Source Cloud marketplaceBedrock, Azure OpenAI, GCP. Bought under an existing cloud contract; regions and model versions often lag the direct API.Parity or noted~ Noted only Consumer subscriptionPro / Plus / Max plans. Buys a product with usage limits, not tokens at a rate.Per month✗ Never Open weights / self-hostLicensed weights you run yourself. No per-token fee; cost becomes your own compute.License✗ Not a rate Gateways / resellersAggregators and third-party inference hosts. Useful as a route, priced at their own margin.Resale✗ Not source ⚠ A SUBSCRIPTION IS NOT A TOKEN PRICE A $20/month plan and a $3-per-million-tokens API rate are different units of a different product, and mixing them is the fastest way to publish a number that means nothing. Subscription figures live on the access card and in the subscription vs API calculator, which compares them on a modelled workload. They never enter the per-token dataset, and they never appear in a pricing table as if they were a rate. Resale prices are never a source. A gateway can name its own margin and change it without the vendor doing anything, so a number read off a reseller tells you about the reseller. We'll name those routes when they matter — MIT-licensed weights served from US/EU regions solve a real data-residency problem — but the price stays theirs, not the vendor's.
Route availability is also dated and specific: which regions a marketplace serves, which model versions it actually carries, whether data stays in-region. Where we can't verify a route's terms first-hand, the card reports what the vendor documents and doesn't dress it up as tested.
§ 04
[H2] Where scores come from.
#score-sourcing We do not run benchmarks ourselves. Every quality score on this site is read off a third-party leaderboard that evaluates models under one standardized harness, and each score carries the leaderboard it came from and the date we read it. We state that plainly because the alternative is worse: a desk this size claiming its own 500-task agent runs would be claiming a compute budget it does not have, and you would have no way to check.
We lean on neutral leaderboards rather than vendor numbers because vendor-published scores are not comparable to each other. Custom scaffolding, tuned prompts and larger context budgets are all normal practice in a launch post, and none of them are labeled. One third-party harness applied to every model is a weaker claim than a self-run — but it is an honest one, and it is the same claim for every model on the board.
SourceWhat we take from itScore cellsStatus llm-stats.comPer-benchmark neutral leaderboards — coding, knowledge, reasoning, math47✓ Primary tbench.aiTerminal-Bench v2 — agentic terminal tasks4✓ Primary Vendor-reportedLast resort when no neutral coverage exists; always carries a visible "vendor-reported" tag0~ Flagged Self-runReserved in the schema; no self-run score is published today0✗ None
Every cell records its value, source, source URL, the configuration the leaderboard used, a provenance tag, and a verification date. A score whose verification date is older than 30 days is re-pulled before it can be published again; it is never quietly re-served as current. All 51 cells on the board today are neutral third-party results.
⚠ AN ABSENT SCORE IS NOT A LOW SCORE Missing results are omitted, never estimated, and coverage is computed from what is actually present. A model becomes rankable only once it carries at least three core benchmarks including one coding and one reasoning or knowledge result — below that it can show scores but takes no rank. Closed frontier models and several Chinese vendors are genuinely under-covered on open leaderboards, and filling those gaps with a plausible-looking number would turn a coverage hole into a finding. § 05
[H2] Benchmark handling.
#benchmarks
We track 8 benchmarks, and they are not all equal. 6 are core: they carry a fixed weight and are the only ones that feed a ranking. The rest are profile-only badges — shown on a model where we have them, never folded into a score. Keeping that line visible matters, because a benchmark quietly promoted from badge to ranking input would change every board on the site without anyone announcing it.
BenchmarkDimensionWeightLeaderboard LiveCodeBench coding 20% llm-stats ↗ SWE-bench Verified agentic-coding 20% llm-stats ↗ Terminal-Bench v2 agentic 15% tbench.ai ↗ MMLU-Pro knowledge 25% llm-stats ↗ GPQA Diamond reasoning 10% llm-stats ↗ AIME 2025/2026 math 10% llm-stats ↗ SWE-bench Pro agentic-coding — profile only scale ↗ Humanity's Last Exam reasoning — profile only artificialanalysis ↗ 4 of the 6 core tests are flagged as saturated — 2 where the field as a whole has topped out and 2 where only the leaders have converged. We still show them, because a saturated test is useful evidence that a cheap model has caught up with an expensive one. We just don't let a one-point lead on a saturated test decide a ranking, since at that end of the scale the gap between two models is noise rather than signal.
We do not publish a page per benchmark, and there is no reproduction script to link to — see § 04 for why. What each score does carry is the leaderboard it came from, the configuration that leaderboard ran, and the date we read it, printed next to the number wherever it appears.
§ 06
[H2] How rankings work.
#rankings
Editorial rankings (Top 10 lists, Best-of picks) follow a strict process: rubric locked → tools scored independently → outliers re-reviewed → final ranking computed. Affiliate relationships exist but are disclosed per-link.
01
[H4] Rubric locked
For each ranking, we define 6-8 criteria + explicit weights (e.g., Aesthetic 25%, Adherence 20%, Cost 15%). Rubric is published before testing starts. Any change to rubric requires version bump. 02
[H4] Pass one — rubric scoring
Founder scores every candidate against the locked rubric, criterion by criterion. Each candidate is scored in isolation; the previous candidate's scores are not visible while scoring the next one. Sponsor status is not visible at this stage. One scorer, one rubric, no shortcuts. 03
[H4] Pass two — checklist sweep
A separate checklist pass verifies the scores against the underlying evidence: each criterion must point at a specific source, screenshot, or test run. Anything not anchored to evidence gets flagged for re-scoring. This catches the "I just felt it" kind of mistake. 04
[H4] Pass three — cool-off review
After a 48-hour cool-off, the founder re-reads the entire ranking from a fresh head. Score adjustments at this stage are rare but logged with the reason. Final 0-100 score is weighted per rubric and summed; ranking is deterministic from the score. If a sponsor lands in #1 by the math, that's the math. One person, three passes, no manual override. § 07
[H2] How head-to-heads are built.
#comparisons
A comparison page — there are 50 of them carries no numbers of its own. Each side declares a reference into the pricing snapshot, and every price, context window and effective-cost figure resolves from that snapshot when the site is built. This is deliberate: a comparison that stores its own copy of a price will eventually contradict the model page it links to, and the reader has no way to tell which one is stale.
01
[H4] Pair chosen
A pair ships only when both sides can be sourced the same way. Where one side has no neutral coverage — IDE plugins with no agent benchmark that measures both, for example — the pair is held rather than published with one side estimated. Page order follows whichever direction people actually search for; the reverse spelling redirects to it instead of splitting the same content across two URLs. 02
[H4] Figures resolve, never typed
Derived rows read the snapshot at build time. If a reference is missing or doesn't resolve, the build fails — it does not fall back to a blank cell or a remembered number. The only hand-written figures are ones the snapshot genuinely can't hold, such as subscription plan prices, and those are marked as authored. 03
[H4] Row winners computed
Per-row wins are arithmetic, not opinion: cheaper wins on a rate, larger wins on context. Where one side publishes no cache or batch price at all, the row shows a dash and awards nobody the win — "no published cache rate" is not "infinitely expensive," and treating absence as a loss would quietly punish vendors for thin documentation. 04
[H4] Gate before publish
A verification script runs over every comparison before it ships. It blocks unfilled placeholders, references that don't resolve, model pairs with no benchmark section, and out-of-rang

Last verified 2026-10-30 LIVE PRICING343 MODELSUSD BASELINEDAILY VERIFIED
[H1] LLM pricing
The independent pricing desk for every frontier LLM API. We pull rates directly from vendor pages, compute blended effective costs (agentic 92/8, cache 82%), and publish the math you can audit. No obscured fine print. Browse all models →Open calculatorCompare models Models tracked 343 across 35 providers Cheapest input $0.0298/M Doubao Seed 2.0 Mini Biggest cache discount −99% MiMo-V2.6-Pro Last refresh 2026-10-30 vendor pages Longest context 1.05M GPT-6 Sol Latest event JAN 01 Scheduled by Google: input r § 01 / EDITORIAL SPOTLIGHTS
[H2] The desk is watching.
All models → ANTHROPIC · MYTHOS
[H3] Claude Fable 5.1 by Anthropic
Flagship model at $10/M input · $50/M output with a 1M window. Prompt caching cuts cached input to $0.25/M — a 98% discount that drops effective cost dramatically for agentic and tool-use workloads. $10/M input$50/M output1M context$0.25/M cached ↗ SAKANA AI · PERFORMANCE-MULTI-AGENT
[H3] Fugu Ultra by Sakana AI
$5 input / $30 output per million tokens. 1M context window. Full pricing hub live. $5/M in$30/M out1M ↗ FIREWORKS AI · FRONTIER
[H3] Ember-1 by Fireworks AI
$3 input / $15 output per million tokens. 1M context window. Full pricing hub live. $3/M in$15/M out1M ↗ GOOGLE DEEPMIND · FRONTIER
[H3] Gemini 4 Argon by Google DeepMind
$2 input / $10 output per million tokens. 1M context window. Full pricing hub live. $2/M in$10/M out1M ↗ OPENAI · FRONTIER
[H3] GPT-6.1 Sol by OpenAI
$2 input / $10 output per million tokens. 1.05M context window. Full pricing hub live. $2/M in$10/M out1.05M ↗ § 02 / PROVIDERS
[H2] Browse by provider.
All models → OPEN-WEIGHT GIANT
[H3] Alibaba (Qwen)
The Qwen (Tongyi Qianwen) model family from Alibaba Cloud's Tongyi Lab. First released in 2023, Qwen is the most-downloaded open-weight model family in the world — most tiers ship under Apache 2.0 on Hugging Face and ModelScope, while the proprietary Max tier is served through Alibaba Cloud's Model Studio. 18 models · from $0.03/M → GLM ARCHITECT
[H3] Zhipu (Z.ai / GLM)
The Beijing lab behind the GLM models, also operating internationally as Z.ai. Spun out of Tsinghua University in 2019, Zhipu ships strong coding/agentic models, multiple free tiers, and open weights — and was the first of China's "Six Tigers" to pursue an IPO. It also sits on the US Entity List. 17 models · from $0.07/M → FRONTIER LAB
[H3] OpenAI
The lab behind ChatGPT and the GPT family. Founded as a nonprofit in 2015 by Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman and others; restructured into a capped-profit LLC in 2019 and finally into a public-benefit corporation on October 28, 2025. Ships the largest commercial model lineup, the largest consumer AI product, and — after the April 2026 funding round — the highest-valued private AI company in history at ~$852B. 16 models · from $0.02/M → FRONTIER LAB
[H3] Anthropic
The AI safety-focused lab behind Claude. Founded by ex-OpenAI researchers in 2021, it ships frontier-grade reasoning models through an API that does not retain prompts by default, and publishes a Responsible Scaling Policy, now in version 3.0. 13 models · from $0.1/M → MEDICAL-AI SPECIALIST
[H3] Baichuan
The Beijing lab behind the Baichuan general models and the medical M-series. Founded in 2023 by Wang Xiaochuan (former CEO of search engine Sogou), Baichuan has pivoted hard toward healthcare AI — its open-source Baichuan-M3 tops several medical benchmarks — while keeping a general-purpose 4-series lineup. 10 models · from $0.0744/M → WECHAT-SCALE LAB
[H3] Tencent (Hunyuan)
The Hunyuan (混元) model family from Tencent — parent of WeChat, QQ, and the world's largest games business. Served through Tencent Cloud and the Yuanbao assistant inside WeChat, Hunyuan reaches over a billion users through China's dominant super-app. 9 models · from $0.0744/M → CHINA CONSUMER LEADER
[H3] ByteDance (Doubao)
The Doubao model family from ByteDance — parent of TikTok and Douyin. Served through the Volcano Ark (火山方舟) platform on ByteDance's Volcano Engine cloud, Doubao powers China's most-used consumer AI app and undercuts most frontier labs on price. 8 models · from $0.0298/M → FRONTIER LAB
[H3] SpaceXSI
Elon Musk's AI lab, founded March 2023 as X.AI Corp. Ships the Grok family — frontier reasoning and coding models trained on the Memphis Colossus supercluster — 200,000 H100s in one interconnected cluster, 150 MW at peak, with a stated roadmap to 1M GPUs. Acquired by SpaceX on Feb 2, 2026 at a $250B valuation ($1.25T combined entity). Renamed from xAI to SpaceXAI in July 2026, when the lab folded fully into SpaceX, and to SpaceXSI (SpaceX Super Intelligence) in October 2026. The Grok brand is unchanged, the API still lives at docs.x.ai, and Grok stays integrated natively into X (Twitter). 8 models · from $1/M → EU FRONTIER LAB
[H3] Mistral AI
The Paris lab behind Le Chat and the open-weight Mistral family. Founded in 2023 by Arthur Mensch with ex-DeepMind and Meta researchers, it ships frontier-grade models with Apache-2.0 open weights for most tiers and EU-native data residency — the leading European alternative to US labs. 7 models · from $0.1/M → FRONTIER LAB
[H3] Google
Google DeepMind is the AI division of Alphabet, formed in April 2023 by merging the original DeepMind (founded London, 2010, acquired by Google in 2014) with Google Brain. CEO Demis Hassabis. Ships the Gemini family — the most multimodal frontier lineup on the market (text + image + audio + video native) — through both the Gemini API and the cloud platform now rebranded as Gemini Enterprise Agent Platform (formerly Vertex AI, renamed April 2026). 4 models · from $0.3/M → NO PER-TOKEN RATE
[H3] NVIDIA
NVIDIA builds the Nemotron open model family and serves it through NIM inference microservices. It is the one major lab on this site that does not sell its models by the token — you evaluate free, self-host under a licence, or buy from a partner. That makes "the price of Nemotron" a different question from the one this site usually answers. 4 models → ENTERPRISE AI
[H3] Cohere
Cohere is an enterprise AI company founded in 2019 in Toronto by Aidan Gomez, Nick Frosst and Ivan Zhang. It sells the North platform, the Command text models, and Embed, Rerank, Parse and Transcribe. Only part of that is priced per token in public: Command R, Command R7B and Aya Expanse. Its newest models — Command A+ and North Mini Code, both Apache-2.0 — are listed as free rather than with a per-token rate, and Command R+ is now priced only for existing customers. 3 models · from $0.0375/M → DIFFUSION LLMs
[H3] Inception (Mercury)
Inception builds diffusion LLMs — models that generate tokens in parallel rather than one after another. The Mercury family is the only architecture of its kind priced in this catalogue, and the company sells it on throughput: over 1,000 tokens per second, or as its own marketing puts it, frontier quality at five times the speed. 3 models · from $0.2/M → MULTI-AGENT MODEL
[H3] Sakana AI
Sakana AI describes Sakana Fugu as a multi-agent system delivered as one model API. Fugu dynamically orchestrates a pool of expert models for coding, reasoning, research, security analysis, and other complex multi-step work. Since September 11, 2026 it comes in two fixed-price tiers: Fugu Ultra for the highest capability and Fugu Max for the lowest cost. 3 models · from $0.95/M → EU / ISRAEL LAB
[H3] AI21 Labs
Jamba is the model family from AI21 Labs, an Israeli AI lab founded in 2017. Jamba pairs a Mamba state-space architecture with Transformer layers for a 256K-token context at a lower cost than comparable long-context models, and ships as Jamba Large and Jamba Mini. 2 models · from $0.2/M → CHINA FRONTIER
[H3] Baidu
ERNIE (Wenxin) is the large-language-model family built by Baidu, China's largest search company, and served through its Qianfan MaaS platform. The lineup runs from the ERNIE 5.x frontier tier down to the cost-efficient ERNIE 4.5 Turbo, all priced per token in Chinese yuan. 2 models · from $0.119/M → COST DISRUPTOR
[H3] DeepSeek
The Chinese lab behind the DeepSeek open-weight models. Spun out of quant hedge fund High-Flyer in 2023 by Liang Wenfeng, it shocked the industry by matching frontier reasoning quality at a fraction of the cost — and shipping the weights under an MIT license. 2 models · from $0.3/M → AGENTIC + MEDIA LAB
[H3] MiniMax
The Shanghai lab behind the MiniMax M3 language models and the Hailuo video/audio generators. Founded in 2021 by Yan Junjie (a former SenseTime research lead), MiniMax pairs cheap long-context multimodal models with consumer media apps — and listed in Hong Kong in January 2026. 2 models · from $0.15/M → OPEN-WEIGHT FRONTIER
[H3] Moonshot (Kimi)
The lab behind Kimi — the long-context assistant — and the open-weight Kimi K models. Founded in 2023 by Yang Zhilin with Tsinghua schoolmates, Moonshot ships 3T-class open-weight models with 1M-token context: Kimi K3 in July 2026, and Kimi K3.1 Pro in October, with its weights promised by November 4. 2 models · from $2/M → MULTIMODAL LAB
[H3] Reka
Reka is a multimodal AI lab in Sunnyvale, California, founded in 2022. Its API now runs as the Reka Gateway: one OpenAI-compatible endpoint that serves two Reka models, Reka Flash 3 and Reka Edge, beside open models from other labs that Reka hosts on its own GPUs. Both Reka models also ship as open weights. The older Reka Core and Reka Flash are no longer in the catalog. 2 models · from $0.1/M → KOREA'S FIRST AI UNICORN
[H3] Upstage
The South Korean lab behind the Solar models, founded in October 2020 by Sung Kim — a former HKUST professor who led Naver's Clova team. Upstage pairs a compact, cost-efficient LLM line with a strong document-AI business, and in 2026 became Korea's first generative-AI unicorn. 2 models · from $0.1/M → CLOUD GIANT
[H3] Amazon
Amazon Nova is the family of first-party foundation models built by Amazon and served on Amazon Bedrock. The current Nova 2 generation replaced the original Nova Micro/Lite/Pro/Premier lineup, led by Nova 2 Lite at a 1M-token context window with on-demand per-token pricing. 1 model · from $0.3/M → WEB-GROUNDED
[H3] Perplexity
Perplexity's API centers on web-grounded answers. On September 27, 2026 it ended its Sonar Chat Completions API, which billed tokens plus a per-request search fee. Its Agent API replaces it: each model bills its own token rates and every tool call - web search, URL fetch, sandbox - is metered separately. Perplexity's own model there is Sonar. 1 model · from $0.25/M → § 02 / HEAD-TO-HEADS
[H2] Most-read comparisons.
All comparisons → 2-WAY Antigravity vs Cursor Antigravity is free and generally available; paid Cursor starts at $8 on its own models, and at $30 for anything else. On Artificial Analysi… Google vs Anysphere↗ 2-WAY ChatGPT Free vs ChatGPT Paid ChatGPT Free gives unlimited text chat on GPT-5.6 Luna but meters every tool — uploads, images, voice, deep research — and may show ads. Pai… OpenAI vs OpenAI↗ 2-WAY ChatGPT Go vs ChatGPT Plus ChatGPT Go ($8/mo) is Free with higher tool limits — more messages with tools, uploads, images and memory — on GPT-5.6 Luna, and it may show… OpenAI vs OpenAI↗ 2-WAY ChatGPT Plus vs ChatGPT Pro ChatGPT Plus ($20/mo) and Pro ($100, $200 or $500) share the apps and the GPT-6 models in Work and Codex. Pro adds Extra High and the Pro mo… OpenAI vs OpenAI↗ 2-WAY ChatGPT Plus vs ChatGPT Team ChatGPT Plus is $20/mo for one person. ChatGPT Team — renamed ChatGPT Business — is $20/user/mo billed annually ($25 monthly, 2-seat minimum… OpenAI vs OpenAI↗ 2-WAY Claude Code vs GitHub Copilot Claude Code is a terminal agent running Anthropic models, and it tops the neutral coding-agent leaderboards. GitHub Copilot is a layer acros… Anthropic vs GitHub↗ § 03 / SCOREBOARD
[H2] Live pricing leaders.
Methodology → Metric What it measures Current leader Top value Models tracked Last updated Cheapest input Lowest $/M input tokens — for high-volume RAG and bulk classification. Doubao Seed 2.0 Mini $0.0298/M 148 10-29 Cheapest output Lowest $/M output tokens — matters most for long-form generation. Ministral 3 3B $0.10/M 148 10-09 Biggest cache discount Largest gap between fresh-input and cached-input pricing — best for agentic workloads. MiMo-V2.6-Pro −99% 105 10-21 Longest context Largest context window — best for big-document analysis at high input cost. GPT-6 Sol 1.05M 152 10-28 Most expensive output Highest $/M output — usually the frontier tier, used selectively. GPT-5.5 Pro $180/M 152 10-28
[H3] The tape doesn't lie.
Every price change, cache rollout, and model release in the LLM cluster — logged within hours, not weeks. → Subscribe to the wire JAN 01 Gemini 3.8 Flash · Scheduled by Google: input rises to $1.50/M, output to $7.50/M, cached input to $0.15/M, and cache storage from $0.50 to $1.00 per 1M tokens per hour. To be re-checked and recorded as fact on the day, not assumed PRICING → DEC 01 Gemini 4 Argon · Scheduled by Google: input rises to $4.00/M, output to $20.00/M, cached input to $0.20/M, and Batch, Flex and Priority double. To be re-checked and recorded as fact on the day, not assumed PRICING → OCT 30 Qwen3.6 27B · Stored cached rate removed: it equalled input, and Alibaba's Context Cache page lists no cache for this model. Input and output unchanged. PRICING → OCT 30 Qwen3.6 35B A3B · Stored cached rate removed: it equalled input, and Alibaba's Context Cache page lists no cache for this model. Input and output unchanged. PRICING → OCT 30 Qwen3.6 Flash · Cached rate corrected from Alibaba's Context Cache page: explicit-cache hits bill at 10% of input, creation at 125%. Input and output unchanged. PRICING → OCT 30 Qwen3.6 Plus · Cached rate added from Alibaba's Context Cache page: explicit-cache hits bill at 10% of input, creation at 125%. Input and output unchanged. PRICING → OCT 30 Solar Pro 2 · Retired by Upstage: end of service reached, solar-pro2-251215 listed as deprecated on October 30, 2026, with the solar-pro4 alias as the upgrade path PRICING → OCT 30 Solar Pro 3 · Retired by Upstage: end of service reached, solar-pro3-260323 listed as deprecated on October 30, 2026, with the solar-pro4 alias as the upgrade path PRICING → § 04 / BEST FOR
[H2] Pick by your workload.
If you need frontier reasoning + tool use with cache enabled… Claude Fable 5.1 See pricing → If you're optimising for cost per million tokens in bulk RAG… Doubao Seed 2.0 Mini See pricing → If long context matters more than price… GPT-6 Sol See pricing → If you want balance of cost and capability for a chatbot… Claude Sonnet 5.5 See pricing → If the workload is simple classification at scale… Claude Haiku 5.5 See pricing → § 05 / FULL DIRECTORY
[H2] All 170 active models.
Methodology → Filter All170 01.AI1 AI21 Labs2 Alibaba (Qwen)18 Amazon1 Anthropic15 Baichuan10 Baidu2 ByteDance (Doubao)8 Cohere3 DeepSeek2 Elorian1 Fi

Last verified 2026-10-30 CompareObjective verdictsLive pricing
[H1] Compare50
Side-by-side AI model head-to-heads — live pricing, neutral benchmarks, and a clear verdict by use case. We call the winner where there is one, and say "tie" where there isn't. Browse all →Model pricing →AI agents →How we compare → Head-to-heads 50 2-way pairs Models compared 18 across all pairs Providers 19 brands covered Benchmarks 6core neutral leaderboards Pricing live from vendor pages Verified 10/30 2026 § 02 / HEAD-TO-HEADS
[H2] Model head-to-heads.
2-WAY Antigravity vs Cursor Antigravity is free and generally available; paid Cursor starts at $8 on its own models, and at $30 for anything else. On Artificial Analysis's Coding… Google vs Anysphere↗ 2-WAY Claude Code vs GitHub Copilot Claude Code is a terminal agent running Anthropic models, and it tops the neutral coding-agent leaderboards. GitHub Copilot is a layer across your edi… Anthropic vs GitHub↗ 2-WAY Claude Sonnet vs Claude Haiku Same family, two different jobs. Claude Haiku is half the price ($1 / $5 vs $2 / $10 per 1M) and the fastest tier on Anthropic's own latency table. Cl… Anthropic vs Anthropic↗ 2-WAY Claude Sonnet vs Claude Opus Same family, two jobs. Claude Sonnet is the 2.5× cheaper workhorse ($2 / $10 vs $5 / $25 per 1M) that handles roughly 90% of coding and chat work. Cla… Anthropic vs Anthropic↗ 2-WAY Claude vs ChatGPT Both are frontier assistants that trade wins. Claude leads real-world coding (SWE-bench Pro 69.2 vs 63.4), long-form writing, and is MCP-native. ChatG… Anthropic vs OpenAI↗ 2-WAY Claude vs Gemini These two split cleanly. Claude Opus 4.8 wins quality — real-world coding (SWE-bench Pro 69.2 vs 55.1), frontier reasoning, long-form writing. Gemini … Anthropic vs Google↗ 2-WAY Cline vs Cursor Two cost models, not two intelligences. Cline is a free Apache-2.0 extension on bring-your-own-key: you pay the provider directly, on any model, with … Cline Bot vs Anysphere↗ 2-WAY Codex vs Claude Code Both are terminal-plus-cloud coding agents that run neck-and-neck. Codex (OpenAI, GPT-5.x-Codex) has more surfaces — CLI, IDE, cloud, GitHub, iOS, Chr… OpenAI vs Anthropic↗ 2-WAY Codex vs Cursor Codex is the stronger agent and Cursor is the cheaper one, by wide margins in both directions. On the Artificial Analysis Coding Agent Index, Codex on… OpenAI vs Anysphere↗ 2-WAY Microsoft Copilot vs ChatGPT This is Microsoft's consumer assistant, not GitHub Copilot. Copilot is free, built into Windows, and its paid path — Microsoft 365 Premium at $19.99/m… Microsoft vs OpenAI↗ 2-WAY Copilot vs Gemini Both sell AI as an entitlement on a productivity subscription, and both charge $19.99 for the main tier. Gemini names its models, ships a real free ti… Microsoft vs Google↗ 2-WAY Cursor vs Claude Code Two different shapes of tool. Cursor is an AI-native IDE that runs Claude, GPT, Gemini or its own Composer; Claude Code is a terminal-native agent on … Anysphere vs Anthropic↗ 2-WAY Cursor vs GitHub Copilot This is GitHub Copilot, the coding tool — not Microsoft's consumer assistant. GitHub Copilot carries Claude and GPT from $10/mo, charges $19 a team se… Anysphere vs GitHub↗ 2-WAY Cursor vs VS Code Cursor is a fork of VS Code, so the editor underneath is the same. What you pay for is the AI layer: a house model at $0.55 per agent task, per-hunk m… Anysphere vs Microsoft↗ 2-WAY DeepSeek vs ChatGPT DeepSeek and ChatGPT answer different questions. DeepSeek V4 Pro is open-weight and cheaper — the API undercuts GPT-5.6 Terra on input and output, mos… DeepSeek vs OpenAI↗ 2-WAY DeepSeek vs Claude DeepSeek and Claude sit at opposite ends of the price/quality curve. Claude Opus 4.8 wins quality — GPQA Diamond 93.6 vs 90.1, SWE-bench Verified 88.6… DeepSeek vs Anthropic↗ 2-WAY DeepSeek vs Gemini DeepSeek sells price and control; Gemini sells reach. DeepSeek V4 Pro is open-weight (MIT) and ~3.4× cheaper on input, ~9× cheaper on output, and you … DeepSeek vs Google↗ 2-WAY Gemini CLI vs Claude Code Gemini CLI is free and open-source — Apache-2.0, roughly 1,000 requests/day on a Google account, backed by Gemini 3.1 Pro's 1M-token, multimodal conte… Google vs Anthropic↗ 2-WAY Gemini Pro vs Gemini Flash Google's tier ladder inverted. Gemini 3.6 Flash (GA, Jul 21) is cheaper ($0.75 / $3.75 vs $2 / $12), about 2× faster, and outscores the older Gemini 3… Google vs Google↗ 2-WAY Gemini vs ChatGPT Gemini and ChatGPT are the two default assistants, and they split the field. Gemini is cheaper on the API ($0.75 / $3.75 vs $2 / $12 per 1M), has a mo… Google vs OpenAI↗ 2-WAY GPT-6 Astra vs Claude Fable 5.1 Two frontier models launched two days apart at the same $10 and $50 per million. Fable 5.1 reads cache at a quarter of OpenAI's rate, so per token it … OpenAI vs Anthropic↗ 2-WAY GPT-6 Astra vs GPT-5.6 Sol OpenAI scaled every cell by the same factor: Astra costs exactly 2.5x Sol on input, cached input, cache writes, output and batch. The composite barely… OpenAI vs OpenAI↗ 2-WAY Grok vs ChatGPT ChatGPT leads the neutral scores — Intelligence Index 52.7 vs 46.5, Terminal-Bench 4.0 59.1% vs 25.8% — and its top model costs five times as much: $1… SpaceXSI vs OpenAI↗ 2-WAY Grok vs Claude Both sides moved in September: Grok 4.7 shipped on Grok 4.6’s card, Claude Opus 5.5 shipped under Opus 5’s. Claude leads every neutral score here — In… SpaceXSI vs Anthropic↗ 2-WAY Grok vs Gemini This pair splits cleanly. Grok 4.7 leads the composite — Intelligence Index 46.5 vs 40.9 — and terminal work, 25.8% vs 19.7% on Terminal-Bench 4.0. Ge… SpaceXSI vs Google↗ 2-WAY Kimi vs Claude Kimi K2.6 is open-weight and roughly 5× cheaper on input, 6× on output than Claude Opus 4.8. Claude wins quality — SWE-bench Verified 88.6 vs 80.2, GP… Moonshot AI vs Anthropic↗ 2-WAY Kimi vs Qwen Two Chinese labs at the frontier, priced in opposite directions. Qwen3.8 Max lists $2/$6 against Kimi K3's $3/$15, so raw output is 2.5× cheaper. But … Moonshot AI vs Alibaba↗ 2-WAY LM Studio vs Ollama Both run open-weight models on your own hardware, and on one machine both can load the same GGUF file. LM Studio is a desktop app with a model browser… LM Studio vs Ollama↗ 2-WAY Ollama vs llama.cpp This is not two competing runtimes. Ollama's own README lists llama.cpp under "Supported backends", so the question is what the wrapper adds. Less tha… Ollama vs ggml-org↗ 2-WAY Perplexity vs ChatGPT Two different tools, not two models. Perplexity is an answer engine: every response is web-grounded with clickable citations by default, and Pro route… Perplexity vs OpenAI↗ 2-WAY Perplexity vs Claude Perplexity and Claude solve different problems. Perplexity is a cited answer engine — real-time web search, inline sources, and Deep Research at $20/m… Perplexity vs Anthropic↗ 2-WAY Perplexity vs Gemini An answer engine against a general assistant. Perplexity grounds every reply in live sources with inline citations by default, and Pro routes your que… Perplexity vs Google↗ 2-WAY Qwen vs ChatGPT Qwen and ChatGPT split the board. Qwen Studio is free with no consumer paid tier, and Qwen3.7 Max is cheaper on output tokens. ChatGPT leads on compos… Alibaba vs OpenAI↗ 2-WAY Qwen vs Claude Qwen and Claude both run 1M-context flagships, and they split cleanly. Claude Opus 4.8 wins quality: SWE-bench Verified 88.6 vs 80.4, GPQA Diamond 93.… Alibaba vs Anthropic↗ 2-WAY Qwen vs DeepSeek China's two biggest model families, and they diverge sharply. Qwen 3.7 Max leads reasoning and knowledge — GPQA 92.4 vs 90.1, MMLU-Pro 89.6 vs 87.5 — … Alibaba vs DeepSeek↗ 2-WAY Replit vs Lovable Both turn a prompt into a running app and bill by credits, but the meters differ. Replit denominates credits in dollars — $25 of credits on a $25 plan… Replit vs Lovable↗ 2-WAY vLLM vs Ollama These are not rivals so much as different jobs. vLLM is a serving engine: continuous batching, PagedAttention, tensor and pipeline parallelism, Linux … vLLM project vs Ollama↗ 2-WAY Windsurf vs Cursor Two VS Code forks that run the same frontier models. Windsurf now ships as Devin Desktop under Cognition, and it wins on cost mechanics — unlimited Ta… Cognition vs Anysphere↗ § 02 / HEAD-TO-HEADS
[H2] Plan & tier comparisons.
PLAN VS PLAN ChatGPT Free vs ChatGPT Paid ChatGPT Free gives unlimited text chat on GPT-5.6 Luna but meters every tool — uploads, images, voice, deep research — and may show ads. Paid (Plus at… OpenAI vs OpenAI↗ PLAN VS PLAN ChatGPT Go vs ChatGPT Plus ChatGPT Go ($8/mo) is Free with higher tool limits — more messages with tools, uploads, images and memory — on GPT-5.6 Luna, and it may show ads. Plus… OpenAI vs OpenAI↗ PLAN VS PLAN ChatGPT Plus vs ChatGPT Pro ChatGPT Plus ($20/mo) and Pro ($100, $200 or $500) share the apps and the GPT-6 models in Work and Codex. Pro adds Extra High and the Pro models in Ch… OpenAI vs OpenAI↗ PLAN VS PLAN ChatGPT Plus vs ChatGPT Team ChatGPT Plus is $20/mo for one person. ChatGPT Team — renamed ChatGPT Business — is $20/user/mo billed annually ($25 monthly, 2-seat minimum). Same 54… OpenAI vs OpenAI↗ PLAN VS PLAN Claude Free vs Claude Pro Since June 30, 2026 Free and Pro run the same default model, Sonnet 5 — so raw answer quality is identical. Pro's $20/mo buys roughly 5× the usage, Op… Anthropic vs Anthropic↗ PLAN VS PLAN Claude Max 5x vs Max 20x Two rungs of one plan. Max 5x is $100 a month, Max 20x is $200 — and across all 53 rows of Anthropic's own plan-comparison table, every other cell is … Anthropic vs Anthropic↗ PLAN VS PLAN Claude Pro vs Claude Max Same Claude, same Opus 4.8 — the only difference is how much you can run. Pro is $20/mo and enough for most; Max is $100 (5×) or $200 (20×) the per-se… Anthropic vs Anthropic↗ PLAN VS PLAN Claude Team vs Claude Enterprise Team is a listed price with usage baked in: $20–$25 a standard seat, $100–$125 premium, minimum two members, and heavy weeks get throttled. Enterprise… Anthropic vs Anthropic↗ PLAN VS PLAN Gemini Free vs Google AI Pro Same models, different headroom. Gemini Free already runs Gemini 3 Pro at $0, but caps you at a 32k context window and standard usage. Google AI Pro (… Google vs Google↗ PLAN VS PLAN Google AI Pro vs Google AI Ultra Same Gemini models, different headroom. Google AI Pro is $19.99/mo — 4x free limits, 1,000 Flow credits, 5 TB. Google AI Ultra starts at $99.99/mo (5x… Google vs Google↗ PLAN VS PLAN Grok Free vs SuperGrok Grok Free ($0) gives limited Chat, Imagine and Grok Build. SuperGrok closed to new customers on October 12, 2026: existing plans continue, and new buy… SpaceXSI vs SpaceXSI↗ PLAN VS PLAN Perplexity Free vs Perplexity Pro Perplexity Free and Pro run the same engine — the gap is volume and access. Free caps you at 3 Pro Searches a day, one Research report a month, and no… Perplexity vs Perplexity↗

Last updated 2026-10-28 / 20 items Journal Independent editorial
[H1] The journal.
20 posts across analysis, opinion, how-to, research, and roundups. 20 Posts tracked 0 New this month 16min avg Avg read time About Independent analysis, opinion, and how-to coverage of LLM tooling -- written for engineers who pay the bill. No vendor talking points, no benchmarks lifted from press kits. Every model price comes from the vendor's own page, verified the week of publication; every cost calculation shows its math. The blog sits next to the price desk and the comparisons -- prose covers the questions a table cannot answer (when does a 90% cache discount stop mattering? which workload class breaks GPT vs Claude pricing parity?). Cornerstone analysis ships first; opinion, tools how-to, research summaries, and roundups will join as the catalogue grows. Read for the math, stay for the disagreements. § Filter by category Category All Analysis (7)OpinionTools & how-to (10)ResearchRoundups (3) Tools Howto
[H2] A 429 is free. The ninth retry is not.
A 429 bills you nothing - the request never reaches the model. The cost arrives when your backoff outlives the cache TTL and the cold retry burns 101x the limit. By Yaroslav V. · Sep 19, 2026 · 17 min read Tools Howto 02
[H2] input_tokens means three different things. Your cost function knows one of them.
Anthropic's input_tokens excludes cached tokens, OpenAI's includes them. Port a cost function between the two and a $0.082 request bills $0.378. Fields, traps, code. openaianthropicgemini Analysis 03
[H2] The list price is not the price your agent pays
The model 5.7% cheaper on the price tag billed an agent session 4.4x higher. The cached-read rate decides an agent bill, and it runs from 2% to 25% of input. prompt-cachingcost-optimizationagents Analysis 04
[H2] Running many agents at once: the tax is linear, the rebate is quadratic
Every extra agent pays for its own cold prefix. But re-reading history grows with the square of turns, so splitting one job across three agents ran 17% cheaper. claude-codeagentssubagents Tools Howto 05
[H2] The 20x in Claude Code caching is real — it's a penalty, not a discount
A cache miss costs exactly 20x the warm read of the same history. What that does to your plan limits, which settings change it, and the number nobody can compute. claude-codeprompt-cachinganthropic Tools Howto 06
[H2] The Batch API discount isn't 50% -- and it isn't always a discount
Batch halves token rates at OpenAI, Anthropic and Google -- but takes only 20% at SpaceXSI. When batch beats a warm cache, and when it costs more. batch-apicost-optimizationapi-pricing Tools Howto 07
[H2] How to actually get prompt cache hits (and why yours are missing)
Prompt caching fails silently -- no error, no warning, just a full-price bill. The floors, the breakpoint rules, and everything that quietly invalidates a cache. prompt-cachingcost-optimizationapi-pricing Analysis 08
[H2] Routing by task complexity: what the 2026 evidence actually shows
Ten routers launched on HN in five weeks. The peer-reviewed 2026 benchmarks say they barely beat a baseline. The arithmetic that decides it, on verified rates. cost-optimizationmodel-routingapi-pricing Tools Howto 09
[H2] The context engineering playbook: which lever, in what order, and what each is worth
Five levers, one deployment order: which context feature to reach for by symptom, what each does to your cache and your bill, and one config that runs them all. context-engineeringagentscost-optimization Analysis 10
[H2] Token savings are not dollar savings: the conversion math
Cutting 63% of an agent's tokens moved its bill 9%. Cutting 1.2% moved it 15%. The arithmetic that converts token reductions into dollars, on verified 2026 rates. cost-optimizationcontext-engineeringapi-pricing Tools Howto 11
[H2] The MCP context tax: how to stop paying for tools your agent never calls
Five MCP servers burn ~55k tokens before you ask anything. Tool search and programmatic tool calling (both now GA) cut that 85%+ — with one caveat that bites. context-engineeringmcptool-use Tools Howto 12
[H2] Compaction: when to summarize an agent's history — and when it backfires
Anthropic's compaction API summarizes an agent's history when it hits a token threshold. How it works, the billing pass you don't see, and when it backfires. context-engineeringcompactionagents

§01
[H2] Where to write
Email Pick the address that matches your inquiry — it reaches the right queue faster. If nothing fits, [email protected] is always safe. Editorial / corrections [email protected] Subject: "Editorial — <page URL>". Reply 48h target. Press / citations [email protected] Subject: "Press — <outlet/podcast>". Include brief upfront. Partnerships [email protected] Sponsorship, data licensing, co-published research. 5-day decision. Everything else [email protected] General questions, feedback, just-saying-hi. We read every email. Before emailing about a potential correction, please check whether the issue is already logged at /corrections/.