Daily drop

Checked claims for October 22, 2026

Top claim

VerifiedProvisionalrelease

Claude Code v2.1.307 adds /claude-api api-version, which moves code to Anthropic API version 2026-10-21. That version removes temperature, top_p and top_k, which now return a 400 error on every model.

How we checked: Read the GitHub release notes and Anthropic's API versions page, which lists 2026-10-21 and the removed parameters.
Why it matters: Code that still sends sampling parameters will break once it moves to the new anthropic-version header. The new command is meant to help with that migration.
Try this: Before changing the anthropic-version header to 2026-10-21, search your code for temperature, top_p and top_k and read the official migration guide.
Source record: Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Anthropicprimary sourceprimary source

VerifiedProvisionalrelease

Anthropic announced Enterprise Frontier Safeguards: free misuse monitoring that keeps activity data in the customer's own cloud storage. It rolls out in phases later this fall, and eligible customers get ZDR on Fable 5 and 5.1 until then.

How we checked · Why it matters · Try this

How we checked: Anthropic's announcement page confirms the free pricing, customer-owned storage, phased rollout and the interim ZDR offer. Not yet generally available.
Why it matters: Regulated teams blocked by 30-day retention on Fable 5 may get a path to frontier models without Anthropic holding their data.
Try this: If 30-day retention kept your organization off Fable 5, request access through the form on the announcement page and ask whether you qualify for interim ZDR.
Source record: Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2027-01-04

Anthropicprimary source

VerifiedProvisionalrelease

IBM released granite-speech-5.0-2b under Apache 2.0. It does ASR in six languages, translates to and from English, and accepts keyword-biasing lists of up to 2,000 terms.

How we checked · Why it matters · Try this

How we checked: The model card confirms Apache 2.0, six ASR languages and the 2,000-term keyword limit. The #2 WER leaderboard rank was not checked.
Why it matters: This is an open speech model you can steer toward product names and acronyms, a common failure point in voice-agent transcription.
Try this: Test granite-speech-5.0-2b from its Hugging Face card on a sample of your own audio, with and without a keyword list of your domain terms.
Source record: Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Hugging Face Blogprimary source

Docs-verifiedProvisionaltool

Claude Code mods are plugins made of JS/TS event handlers that run inside Claude Code. Six are built in, including you-should-know, which is off by default. Mods are not sandboxed.

How we checked · Why it matters · Try this

How we checked: The Claude Code mods overview docs confirm the definition, the six built-ins, that you-should-know is off by default, and that mods are not sandboxed.
Why it matters: You can reshape your agent's interface and intercept its tool calls. Any third-party mod runs with your full permissions and can see your secrets.
Try this: Before installing a third-party mod, run claude plugin validate on its directory and read the hooks: and calls: lines.
Source record: Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Latent Space (swyx)primary source

Docs-verifiedProvisionaltool

NVIDIA released Kumo Relational, a pretrained model that makes predictions straight from multi-table databases using in-context examples, with no task-specific training. Code is Apache 2.0 and weights use OpenMDW 1.1.

How we checked · Why it matters · Try this

How we checked: The model card and GitHub repo exist with the stated licenses and library. The claim of beating supervised baselines on 41 tasks is NVIDIA's own (vendor claim).
Why it matters: This could replace feature engineering plus XGBoost for churn or demand prediction, and agents can call it through a query language.
Try this: Read the quickstart notebook in NVIDIA/structured-data-models-client and compare the model against your current baseline on one held-out task.
Source record: Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Hugging Face Blogprimary sourceprimary source

Docs-verifiedProvisionaltool

IBM published granite-timeseries-ensemble-r1, an Apache 2.0 recipe that pools four Granite time-series models into one zero-shot forecaster returning median and quantile forecasts, with no fitted weights.

How we checked · Why it matters · Try this

How we checked: The model card confirms the recipe, license and four members. The card says granite-tsfm>=0.3.10 and the blog says >=0.3.11. GIFT-Eval figures differ by snapshot date.
Why it matters: It is a reasonable default when you don't know which pretrained forecaster suits your data. IBM itself says the accuracy gain over the best member is small.
Try this: Backtest the ensemble against PatchTST-FM-r2 alone on your own historical windows, and check 0.1 to 0.9 interval coverage.
Source record: Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Hugging Face Blogprimary source

Vendor claimProvisionalresearch

Anthropic reports that open-weight GLM-5.3 built end-to-end V8 exploits in 50 of 410 attempts (Mythos Preview: 56). Simple bypasses got it to engage with malicious requests 64% to 100% of the time.

How we checked · Why it matters · Try this

How we checked: The GLM-5.3 open-weight model card exists on Hugging Face. The exploit and bypass numbers are Anthropic's own evaluations, not replicated.
Why it matters: Strong exploit-building ability is now downloadable with weak refusals. Defenders should assume attackers have it.
Try this: If you self-host open-weight models in agent pipelines, do not rely on the model's own refusals. Enforce sandboxing and tool permissions outside the model.
Source record: Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Anthropicprimary sourceprimary source

Vendor claimProvisionalresearch

Anthropic says Fable 5.1 in Claude Science computed the nine-loop six-particle amplitude in planar N=4 super-Yang-Mills for about $1,000 to $2,000. Physicist Lance Dixon validated the result.

How we checked · Why it matters

How we checked: Read Anthropic's guest post. Dixon's validation is stated there. An arXiv search found no independent paper yet.
Why it matters: It shows a long-running agent harness completing a fragile frontier computation with little supervision beyond "keep going".
Source record: Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Anthropicprimary source

Vendor claimProvisionaltechnique

Dharma-AI says that rebuilding a Qwen3-4B grader's vocabulary by token cost cut average output from 350 to 101 tokens and measured inference cost by 77%, while its quality score rose from 0.703 to 0.810.

How we checked · Why it matters · Try this

How we checked: The authors' own blog post. The VocabX paper is unpublished and under review, so the results are unreplicated.
Why it matters: For narrow, high-volume fine-tuned models, the tokenizer is a cost lever. Shrinking output tokens saves several times more than shrinking input.
Try this: Measure what share of your serving cost comes from generated versus input tokens before choosing which side to compress.
Source record: Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

Hugging Face Blogprimary source

Seen but not verified (don't act on these)

Not verifiedProvisionalrelease

OpenAI says its frontier models and Codex workflows are coming to Google Cloud through Model Garden and Google Cloud Marketplace.

How we checked · Why it matters

How we checked: OpenAI's feed confirms the announcement exists. The article returned 403, and no Google Cloud docs page confirming availability was found.
Why it matters: Teams standardized on Google Cloud might use OpenAI models under existing billing and procurement. Timing and model list are unconfirmed.
Source record: Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-05

OpenAI (and Sam Altman's blog)primary source

Today's list · All days · latest.md