Corrections log

Was212 pricing pages said "Pricing pulled daily from" the vendor's site in the byline, 109 model pages ended their header with "Pulled directly from {vendor} daily", the /llm-cost/ index described itself as "Updated daily from vendor pages", and twelve provider hubs added a line such as "All OpenAI models tested". None of it was true. Nothing on this site reads vendor prices on a schedule - prices are checked by hand and each page is dated - and the only daily job looks for newly released model ids without reading any price. Across current models the median price was 18 days old and the oldest 93. The date inside the byline was also a hand-kept copy that disagreed with the same page's header on 38 pages. GLM-5 Turbo's page made the point on its own: it promised daily pulls from docs.z.ai for two weeks after Z.ai had removed the model from that page.

NowThe byline now states only what the data shows, and is generated rather than written: "Prices checked by hand against {the vendor page the price was read from} - Last verified {the page's own date}", so it cannot claim a cadence and cannot disagree with the header. The header sentences say "Checked by hand against {vendor}", the /llm-cost/ description says the same, and its answer to "How often are prices verified?" is computed from the catalogue - how many current prices there are, and the dates by which half of them and all of them were last checked - instead of promising a cycle. The build now fails if a pricing page claims a daily cadence again.

Flagged by Internal audit.

WasGLM-5 Turbo was shown as a current model at $1.20/M input, $0.24/M cached and $4/M output, with a calculator, cost scenarios and a place in four other Zhipu pages' comparison shelves recommending it as the latency option. Z.ai had already taken it off the rate card: it is on the Internet Archive's August 24, 2026 capture of docs.z.ai/guides/overview/pricing and gone from the August 29 one, with no retirement notice on any first-party surface. We read that same table on August 27 while adding GLM-5.3-Flash and did not notice, because the check compared the rows it found against the snapshot and had nothing to say about a row that had vanished. The four cost scenarios on the page were also computed without the 82% cache assumption every other page uses, overstating each task by roughly 80%.

NowThe row is marked deprecated with its price frozen at the last published value, and the page now opens by saying the model is off the rate card, dates the window in which it disappeared, and points to GLM-5.2 and GLM-5.3-Flash instead. Because we hold no Z.ai key, the page does not claim the endpoint is dead - only that the vendor stopped listing and pricing it, which is what the evidence shows. The four shelves that recommended it now say it was delisted, and the scenario costs are recomputed on the standard blend ($0.021 per coding task rather than $0.038).

Flagged by Internal audit.

WasEvery Upstage page said Solar Pro 4's 90% launch discount ended on September 10, 2026 and that list pricing - $0.30/$0.06/$1.20 - applied from September 11. That is what Upstage's pricing page said when we recorded it, word for word, and the Internet Archive's August 17 capture confirms it. Upstage has since added a second discount phase without changing the version stamp on the script that renders the schedule, so the claim was about to become false: readers planning September and October spend from these pages would have budgeted three times what they will be charged.

NowThe discount is a ladder, not a single phase: 90% off ($0.03/$0.006/$0.12) through September 10, then 70% off ($0.09/$0.018/$0.36) from September 11 to October 10, then list from October 11. Every page now prints both steps. One editorial claim inverted with it - Solar Pro 4 was said to become twice the price of Solar Pro 3 on September 11, and in fact it stays cheaper on input and output until October 11, while its cache-hit rate ($0.018) does cross above Pro 3's ($0.015) on September 11.

Flagged by Internal audit.

WasGPT-4o was listed at $5/M input, $15/M output and $2.50/M cached, and marked deprecated with an October 23, 2026 shutdown. Every one of those figures was double the real rate, and the whole page was framed as a pricing archive for a dead model. Both errors have one cause: the row described the frozen gpt-4o-2024-05-13 snapshot while carrying the name, slug and page of the plain gpt-4o alias. The old note admitted the source - "pricing widely cited from OpenAI launch announcement" - a launch post from May 2024, never re-read against the price table afterwards. Anyone budgeting a GPT-4o workload from this page overstated it by 2x, and anyone reading it for status was told to migrate off a model OpenAI still sells.

NowGPT-4o now shows the alias's own card: $2.50/M input, $1.25/M cached, $10/M output, plus the Batch ($1.25/$5) and Fast ($4.25/$17) rows, with an effective blended rate of $2.16 rather than $3.91. The cut is first-party and dated - OpenAI's Structured Outputs announcement of August 6, 2024 says switching to gpt-4o-2024-08-06 saves "50% on inputs ($2.50/1M input tokens) and 33% on outputs ($10.00/1M output tokens)" - so the price history now shows the step down instead of a flat line that never happened. Status is legacy, not deprecated: the October 23 retirement names gpt-4o-2024-05-13 alone, and OpenAI's retirement table lists aliases explicitly where they are covered (it writes "gpt-4-0613 | gpt-4" and "gpt-4-turbo | gpt-4-turbo-2024-04-09"), while the gpt-4o row names none. One consequence is worth stating because the page had it backwards: at $30/M output, GPT-5.5 - the model usually named as the GPT-4o replacement - costs 1.7x more than GPT-4o on our blend, not less.

Flagged by Internal audit.

WasFor three days the site said GPT-6 Astra was sold by approval only: an APPROVAL ONLY badge on the home board, RESTRICTED - NOT PUBLIC on the OpenAI hub, "you cannot buy it yet" on the model page, and a whole scored Availability row against it in both comparisons - one of which told readers to build on GPT-5.6 Sol and re-run their evals when the gate opened. That was accurate on September 3, when Astra shipped to Trusted Access Program enterprises. OpenAI opened general access within three days and announced nothing, so the claim rotted in place while every price on the page stayed right.

NowAstra is recorded as generally available, verified September 6 against four independent tells on OpenAI's own docs: rate limits now published across usage tiers 1 to 5 (an approval-gated model has no tier ladder), the models index recommending Astra to any developer unsure where to start, no gating language anywhere in the API changelog or the migration guide, and OpenRouter reselling it since September 4. The rates never changed. This is the second availability claim to go stale in four days - the September 3 entry below is the first - and the class is the same both times: no gate checks a sentence about who may buy, only the numbers. What did work is the build assertion added on September 4, which refused to render the home page rather than let the badge outlive the gate it described.

Flagged by Internal audit.

WasThe hub's "which model should I pick" cards each opened with "If you have preview access and need..." for all three GPT-5.6 tiers. That was true when the cards were written, in the limited-preview window that began June 26, 2026, but the family went generally available across ChatGPT, Codex and the API on July 9 - 56 days before this correction. A reader choosing a model would have read the site's own recommendation for OpenAI's current frontier tier as unavailable to them, and routed to an older, more expensive model instead.

NowThe three cards now describe what each tier is for, with no access precondition, because there is none. Every price on the hub was correct throughout and no calculator, scenario or effective rate was affected - what went stale was a sentence about availability, which is a class that rots more quietly than a price does, since no gate checks it. Found by the intake sweep for GPT-6 Astra, which re-reads the vendor's whole card and then greps our own surfaces for access-status claims. Astra itself is the mirror image: genuinely gated today, and recorded as such.

Flagged by Internal audit.

WasHy3 was listed at $0.141/M input, $0.563/M output and $0.035/M cached input. Those figures were not Tencent's: we produced them by converting the CN-region CNY card (Y1.00 / Y4.00 / Y0.25) at the Y7.10 rate this catalogue used until that day. Tencent Cloud International publishes its own USD price list for the same model at $0.132 / $0.528 / $0.033, so every rate we showed was about 6.8% too high, and so was the effective blended rate and each of the four scenario costs built on it. A reader budgeting from this page would have overstated their Hy3 spend.

NowAll Hy3 rates now come from Tencent Cloud International's USD list, the same source as the new Hy4 Preview row. The two lists are not one number in two currencies: Hy3's USD implies about Y7.58 to the dollar and Hy4 Preview's implies Y7.19, so no single conversion reproduces them and the published USD has to be read rather than derived. Tencent's prices did not change - the price history on that page is deliberately flat, because what moved was our reading of it. This is the same failure we recorded for Alibaba in July, when converting a CN rate card put fourteen prices above the vendor's own International list; the rule that came out of it had not been carried across to Tencent.

Flagged by Internal audit.

WasThe post, published 2026-08-23, said that subagents get the five-minute cache TTL even on a subscription "with no flag on a subscription to widen it", and presented ENABLE_PROMPT_CACHING_1H and FORCE_PROMPT_CACHING_5M as the only TTL controls. Anthropic has since split Claude Code's requests into two TTL buckets - the main conversation, and everything else (subagents, workflows, in-process teammates, forks, compaction, session titles) - and given each bucket its own control. A reader running fan-outs on a subscription would have concluded that the shortest cache lifetime in the product was not adjustable, when it is.

NowThe post now describes both buckets and all four controls: promptCacheTtl and CLAUDE_CODE_PROMPT_CACHE_TTL for the main conversation, subagentPromptCacheTtl and CLAUDE_CODE_SUBAGENT_PROMPT_CACHE_TTL for everything else, each taking 5m or 1h, with FORCE_PROMPT_CACHING_5M overriding both and ENABLE_PROMPT_CACHING_1H now below both per-bucket controls in the precedence order. Anthropic's documentation puts the requirement at Claude Code v2.1.242 or later; the vendor's changelog lists both settings as landing in v2.1.243, after the post was published. No price, calculation or table was affected - the arithmetic in that post is unchanged and still checks against the snapshot. Caught while researching the follow-up piece on parallel agents, which re-read the same documentation page four days later and found it restructured.

Flagged by Internal audit.

WasBoth models were recorded as text-only, so the provider hub's lineup printed "text-only" beside each of them. Alibaba documents both as accepting image input: its Model Studio visual-understanding page names Qwen3.7-Flash and Qwen3.6-Plus among the models that take up to 256 images per request. The field is not cosmetic - it is the only signal on the hub that tells a reader whether a model can be sent a screenshot, and a team filtering for vision would have skipped two models that do it.

NowBoth now record text, vision and code. Per-token prices were correct throughout and are unchanged; no calculator, scenario or effective-rate figure was affected. Found by the control step of our model-intake process, which re-reads a vendor's whole card when a new model from that vendor is added - the same step that caught the Grok modality error in August. This time the check was widened from prices to the vendor's capability documentation, because a price has a shelf of gates behind it and a modality has none.

Flagged by Internal audit.

WasDeepSeek replaced its flat rate card with peak/off-peak billing on August 16, 2026. The stored rates were updated that day and every calculator, tile and scenario followed, but six pieces of prose did not. The provider hub's headline statistic read "Cheapest tier $0.14/M" for a model that now costs $0.44 at peak and $0.22 off-peak; a hub answer compared V4 Pro to GPT-5.5 using $0.87 output when the rate was $3.96, and quoted cache-hit input at $0.0036 instead of $0.044; the hub answered the question "Is there an off-peak discount?" with "No discount" on the very day off-peak became the billing model; and the V4 Flash page still said we had not pre-applied the change. Two further answers were stale in a different direction: both model pages said the Responses API was unsupported on V4 Pro with support "coming in early August 2026", and both repeated a vendor warning about an unspecified future price rise that had already arrived and been removed from the page.

NowAll six now state the live card: $0.44/$0.014/$1.32 per million at peak on V4 Flash, $1.32/$0.044/$3.96 on V4 Pro, and exactly half of each off-peak, which covers 17 hours of a weekday and all weekend. The Responses API is supported on all three DeepSeek models as of today; the remaining difference is concurrency, 500 on Pro against 2500 on Flash. No stored price was ever wrong and no calculator was affected - the defect was confined to hand-written prose and to one hand-written statistic. To stop that statistic going stale again, the hub's model count, cheapest tier and context window are now computed from the snapshot at build time rather than typed in.

Flagged by Internal audit.

WasGrok 4.20 Multi-Agent was published with a 2M-token context window and described as "the only Grok variant with 2M context" — a claim repeated in the model page subtitle and hero tag, the provider hub's headline "Max context" statistic, a use-case pick for "the longest context window in the Grok line", a competitive takeaway against Gemini 2.5 Pro, and the sibling tables on two other Grok pages. SpaceXAI has never published that figure. The same rows also recorded the model as text-only, along with Grok 4.3 and both dated Grok 4.20 snapshots.

NowThe window is 1M. All three SpaceXAI surfaces agree — the pricing table, the model page spec block and the raw page source all state 1,000,000 tokens, and the page contains no 2,000,000 anywhere. Input modalities are text + image for the whole Grok 4.20 family and for Grok 4.3, not text-only. Per-token prices were correct throughout and are unchanged, so calculators, cost scenarios and effective-rate figures were never affected; the error was confined to the context and modality claims. Found by the routine control step in our model-intake process, which re-reads every existing row for a vendor whenever a new model from that vendor is added.

Flagged by Internal audit.

WasEight surfaces stated that Claude Sonnet 5 would move from $2/$10 to $3/$15 per million tokens on September 1, 2026, when its introductory pricing expired. That was true when we published it — Anthropic's pricing page carried the schedule and the September rates. We built on it: the Sonnet-vs-Haiku comparison awarded Haiku a 'price stability' verdict on the strength of the coming rise, one pricing row was devoted to the September rate card, and two blog posts used the date as a worked example of a bill changing without a code change.

NowAnthropic cancelled the increase. Its pricing page now states that $2/$10 'is now the standard price' and that 'the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.' The stored rate was always $2/$10 and never wrong; what was wrong was the forecast attached to it. The scheduled-change block, the pre-computed September rates and the verdict built on them have been removed, and both blog posts carry a dated update rather than a silent edit.

Flagged by Internal audit.

WasAll three GPT-5.6 pages published long-context rates at half the real price. OpenAI's pricing page repeats the same table once per service tier, and at intake on July 10 we read the Batch table's long-context columns as if they were Standard. Sol showed $5/$0.50/$22.50 against a real $10/$1/$45, and Terra and Luna were understated the same way. The figures appeared both in a quote-board footnote and in a dedicated FAQ answer on each page, so a reader sizing a >272K-token request would have budgeted half of what OpenAI charges.

NowLong-context rates are now Sol $10/$1/$45, Terra $4/$0.40/$18 and Luna $0.40/$0.04/$1.80 — the vendor's long-context tier is double the standard input and cached-input rate and 1.5x the standard output rate, which every other OpenAI model in our catalogue already followed. The short-context rates that drive the calculators and every comparison were correct throughout; only the long-context tier was wrong.

Flagged by Internal audit.

WasPrice displays rounded every rate to two decimals, which is too coarse for a catalogue spanning $0.01 to $150 per million tokens. 115 of 770 displayed rates sat more than 1% from the real figure and 21 were more than 5% out — a $0.015 rate read as "$0.01", 33% low. Separately, provider-hub lineup cards rounded anything above $1 to whole dollars, so 198 rates across 132 models were wrong on those cards: Gemini 3.6 Flash's real $1.50 input showed as "$2". Errors ran in both directions, and the visible price disagreed with the exact figure we publish in structured data.

NowRates now render to three significant figures with a two-decimal floor, so $0.435 reads as $0.435 and $1.50 reads as $1.50 while $15.00 keeps its column alignment. Worst-case display error drops from 33% to 0.38% and nothing exceeds 1%. Whole-dollar rounding on lineup cards is gone. The pricing snapshot, calculators and structured data always held the exact values — only the rendering was lossy — and 38 authored page figures that had been written pre-rounded to two decimals were realigned to the snapshot at the same time.

Flagged by Internal audit.

WasThe DeepSeek hub reported a price change that never happened. A dated timeline entry, the V4 Pro lineup card, the lab background and an FAQ all stated that the 75% launch promo ended on May 31, 2026 and that V4 Pro had reverted to $1.74/M input and $3.48/M output — four times the real price. DeepSeek announced that expiry but never applied it. The reversion was written up in advance as a settled fact, so when the vendor let the date pass, the page kept asserting a list price the vendor had never charged.

NowV4 Pro lists at $0.435/M input and $0.87/M output, re-verified against DeepSeek's pricing page on August 2, 2026. The hub now records the expiry date passing without a reprice and treats the discounted figure as the standard rate. The underlying pricing snapshot always held the correct rates, so calculators and model pages were unaffected — the wrong numbers were confined to the hub's prose.

Flagged by Internal audit.

WasThe comparison stated that Cursor has no JetBrains support at all — "Cursor has none", a ✗ in the feature matrix, and a whole scenario advising JetBrains teams away from it on that basis.

NowCursor does reach JetBrains, through the Agent Client Protocol: the agent reads files and runs terminal commands inside IntelliJ, PyCharm and WebStorm. It is weaker than a native plugin — it needs a paid Cursor plan plus the JetBrains AI Assistant plugin 2025.1+, and Tab completion stays in the Cursor app — so Windsurf still wins that row, but the page now says what the limitation actually is instead of denying the feature exists.

Flagged by Internal audit.

WasFourteen Qwen models carried prices converted from Alibaba's China-region CNY rate card — Qwen3.7 Max, for example, at $2.7707/M input and $8.3119/M output. Alibaba publishes its own International (Singapore) list directly in US dollars, and those figures are lower, so every affected rate, calculator result, cost scenario, sibling shelf and comparison page overstated the real price by roughly 8-11%.

NowAll fourteen now carry the vendor's published International USD list, read from the English pricing page: Qwen3.7 Max is $2.50/M input and $7.50/M output. Calculators, price tiles, cost scenarios, price history and every sibling shelf were recomputed to match, and the currency-conversion fields were removed so nothing implies a derived figure. The model-intake note now states that the CN rate card must not be converted when an International USD list exists.

Flagged by Internal audit.

WasThe methodology claimed we run all 500 SWE-bench Verified tasks ourselves on a clean box at ~$170–500 per agent, and that our test harness and result spreadsheets are public. We do none of that — every benchmark score on the site is read off a third-party leaderboard.

NowBoth sections were retracted and rewritten to state plainly that we do not run benchmarks ourselves, that each score carries the leaderboard it came from and the date we read it, and that the figures come from neutral third-party harnesses. The false per-agent cost table and 'reproduced by us' claims are gone.

Flagged by Internal audit.

WasEffective and raw rates in the sibling-comparison tables had gone stale after vendor repricings that never propagated: Grok 4.3 overstated 142%, Gemini 3.5 Flash understated roughly 6×, Mistral Small 4 wrong on 30 pages, and Claude Mythos 5 shown at $13.20 against a real $6.82. The same model could read one effective rate on its own page and a different one in a comparison (Opus 4.8: $3.21 vs $3.41).

NowAll effective-blend and raw rates are now derived from the central pricing snapshot at build time using one shared formula, so a model reads the same number everywhere and cannot drift again. 927 stored values across 188 files were corrected, and a publish gate now fails the build on any disagreement.

Flagged by Internal audit.

WasThe cache-write premium (1.25× input) was recorded only for Anthropic models, so the effective-cost comparison charged Claude for a cost it did not charge GPT — which levies the same premium on GPT-5.6. The page overstated the Claude-vs-GPT cost gap by about 6%.

NowCache-write billing was audited across 16 providers and the GPT-5.6 rates filled in, so both sides are charged on the same basis. Where a vendor genuinely publishes no cache-write rate, that absence is now recorded explicitly rather than read as zero.

Flagged by Internal audit.

WasSub-cent rates such as DeepSeek V4 Flash cached input ($0.0028/M) rendered as '$0.00' across price tables — on a price desk that reads as free — and, worse, were published to Google in schema.org offers as a $0.00 (free) price.

NowA shared sub-cent formatter now shows the real figure to four significant digits, genuine free tiers read 'Free', and schema prices escalate precision so no real rate serialises as $0.00. Zero all-zero dollar strings remain in the build.

Flagged by Internal audit.

WasGoogle Search Console flagged merchant-listing errors on the comparison pages: the structured-data offers were malformed, so the pages' pricing could be misrepresented in Google's shopping surfaces.

NowThe offer markup was corrected to a valid digital-product shape across the comparison pages and the fix verified live; the listings validate cleanly.

Flagged by Google Search Console.