Corrections
When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.
321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.
- editorial
The Life Sciences Verification Program opened to applications from teams and institutions on 17 September 2026 (no longer only invite-only)
- editorial
The Sonnet 5 increase to $3/$15 was cancelled; $2/$10 became the standard price
- editorial
Claude Fable 5 is no longer the current flagship (followed by Fable 5.1 and Opus 5.5). Also: NYT compute figures attributed to investors' expectations; biomolecular speedup precision caveat
- editorial
Opus 4.8 is no longer the newest Opus nor the top Opus SWE-bench Verified score; fast-mode availability updated to current pricing docs (Opus 4.7 fast mode now errors, Opus 4.6 runs at standard rates). Also: tense and count fixes, Terminal-Bench effort note
- editorial
Ablation LOWERS the Llama Guard 2 safety score (the page had it inverted); the paper's limitation concerns untested, larger and proprietary models, not constitutional AI or output classifiers. Also: ORTHO vs GCG comparison, safety-implications paragraph now reflects the paper, validation split, unsourced causal claims removed, FailSpy's first abliterated repo (6 May 2024)
- editorial
All nine implementation partners were named at launch, not added later; adds the 2026 protein design campaigns, the Adaptyv Bio competition, the Life Sciences Verification Program and Anthropic's wet lab
- editorial
Qwen's Qwen-Image-Bench chart covers 29 models, not 26
- editorial
Comparison-model scores in Xiaomi's table are not all Xiaomi's own measurements (several match OpenAI's GPT-6 Astra launch table, and two 'Claude Fable 5' exploit cells match published Claude Mythos 5 results); Artificial Analysis chart lists eight proprietary models ahead, not five. Also: comparison-class definition, Lean case-study wording
- editorial
Speculative decoding speedup is LM Studio's 'up to 1.5x-3x' (not an unsourced 20-50 percent); LM Link comparison row now says v0.4.5+. Also: Splash backend noted in lead and architecture; 0.4.24 changelog cited; Bionic skills and data-retention wording tightened
- editorial
Zhijian Liu is described as a co-founder of Inco AI in NVIDIA's June 2026 DFlash post author bio; the z-lab DFlash 2 drafter repos predate the incoai ones (not mirrors). Also: dated Artificial Analysis provider-speed note
- editorial
The 'tens of thousands of accelerators' wording was removed from Z.ai's launch post by Aug 27; now attributed to the archived version. Also: add Inference infrastructure section from Z.ai's Sep 17 post (GLM-5.3 Infra Agent, dense feedback, 100,000+ accelerators, 62T-token Ox Alpha claim)
- editorial
Xiaomi's model card shows Claude Opus 4.6 at 57.3 on SWE-bench Pro (ahead of V2.5-Pro's 57.2), and HLE is 48.0 with tools / 34.0 without; token-efficiency claim now attributed. Also: successor note (MiMo-V2.6), release dates, removed first-person voice
- editorial
Luo Fuli joined Xiaomi on 12 Nov 2025 (not 'expected to join'); MiMo-7B weights are MIT-licensed on Hugging Face. Also: add MiMo-V2-Pro/Omni, V2.5, V2.6, Reuters 2026 AI budget
- editorial
Sonnet 5 $2/$10 became the standard price (the $3/$15 increase was cancelled). Also: add Claude Opus 5, Fable 5.1/Mythos 5.1 and Opus 5.5 to timeline, milestones, platforms and pricing
- editorial
The Sonnet 5 $3/$15 price increase was cancelled ($2/$10 is standard). Also: add Claude Opus 5 / Fable 5.1 / Opus 5.5 releases, Sept 2026 biomolecular work, Adaptyv competition, wet lab, NYT compute report
- editorial
Abliteration predates the arXiv paper (FailSpy coined it May 2024 after the April 2024 preview post); update Heretic count; link new abliteration and Heretic pages
- editorial
Removed a quote attributed to Anthropic's documentation that no version of the docs contains (replaced with the docs' actual wording); OpenAI does guarantee schema key ordering per its own guide; Anthropic does not support recursive schemas and Gemini added $ref recursion in November 2025; Bedrock availability and beta status per Anthropic's docs and release notes; XGrammar adoption per the vLLM and MLC blogs; SGLang 2x/2.5x/4% figures cited to the LMSYS blog (new ref 40) and an unsourced 30-50% claim removed; Jev Noul answers carry no confidence score; unsourced Guidance 50% claim replaced; three red links removed
- editorial
The API lineage was wrong (text-davinci-001/002 were FeedME models and the paper's PPO recipe first shipped as text-davinci-003; the 2022 API InstructGPT models used a slightly different procedure per OpenAI's blog and model card); two Wired citations that never existed replaced with Leike's own post and the NIST announcement; Almeida is one of the paper's nine asterisked primary authors; the 1.6% compute ratio is computed here, not reported by the paper; SLiC-HF dated 2023
- editorial
On TypeSafe's customer-service workflow three models score above Jev (GPT-5.6 Sol, DeepSeek V4-Flash and V4-Pro), not two; the evals site's 'every model is more accurate, cheaper and faster in the workflow' claim is an average across four tasks with per-workflow exceptions. Minor: 11 Sep post was four days before launch; 150 ms appears in TypeSafe's own docs; Noul answers carry no confidence score
- editorial
In the GPT-4 report, 'Instruction following and API evals' sits under Evaluation & analysis, not Reinforcement Learning & Alignment. Minor: IMO record cited via the team results page; TWIML episode recorded in September, published October 2016
- editorial
GPQA-Diamond, HLE and HMMT 2025 are not the three largest V3.2-Exp drops (Aider-Polyglot fell 1.6 points, twice GPQA's); the report explains only those three. Minor: release-note quote restored verbatim ('DSA'); 'for the first time' quote sourced to the release README; selection-overhead and 'shift' wording attributed; quality claim attributed to DeepSeek's evaluations; dead ReLU link
- editorial
Mixtral 8x7B uses fully dense 32K attention, not sliding-window (its technical report and config); Claude 3.5 Sonnet launched with 200K context, not 128K, and 200K was Claude 3's standard window. Minor: Sparse Transformer fixed pattern attends to the last c positions (not one element); Longformer window schedule per the paper; Atri Rudra is at Buffalo; BigBird path-length wording; NSA authors are Yuan et al.; DSA introduced with V3.2-Exp (Sept 2025); DeepSpeed row dated and cited; Block-Sparse-Attention row per README; 'among the first'; 100,000 squared equals 10 billion; four red links unlinked
- editorial
The paper does not say vLLM or SGLang run out of memory; under its matched on-GPU-cache protocol the full-attention engines run a 16x smaller real batch (only the Hugging Face rows hit OOM). Minor: which studies used 20 sampled HELMET instances; HF repository creation vs modification dates; dated 'Preprint' comment; 16:1 head ratio applies to the 8B model; NOSA-8B compression window discrepancy with the SparDA paper flagged
- editorial
NVIDIA Dynamo's offloading tutorial lists four backends including its own KVBM (the recommended vLLM starting point), not three third-party alternatives; FlexKV tiers are host/SSD/cloud storage and its vendor attribution removed. Minor: NOSA throughput was measured on the 8B model only; NOSA date range; DeepSeek quote wording; HiCache layout wording
- editorial
Browser Use no longer uses Playwright (drives Chrome over CDP; cdp-use dependency), Stagehand SDKs are TypeScript/Python/Go not C#; Nova Act model claim reduced to what the launch post says and citation repointed to an archived copy; unsupported Skyvern WebVoyager score and Comet Android/iOS dates removed; BrowserSkill tab-borrowing wording; playwright red links unlinked