Daily drop
Checked claims for October 16, 2026
Top claim
VerifiedProvisionalrelease
Fast mode for Claude Opus 5.5 no longer needs a beta header or access request. It costs $8 input and $40 output per million tokens.
- How we checked
- Anthropic fast mode docs confirm no beta header for Opus 5.5 and $8/$40 pricing; 2.5x speed is Anthropic's own figure
- Why it matters
- Agent loops slowed down by output can opt in per request at a known premium price (per Anthropic's docs, not on the Batch API or Claude Platform on AWS). Anthropic's speedup figure is its own claim.
- Try this
- Fast mode bills at premium rates and switching speed invalidates the prompt cache, so try it on one existing Opus 5.5 request: add speed: "fast" and check usage.speed in the response.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Latent Space (swyx)primary source
VerifiedProvisionalrelease
Claude Code v2.1.302 adds a notice saying standalone Claude Design at claude.ai/design will close by December 14, 2026.
How we checked · Why it matters · Try this
- How we checked
- Anthropic's own release notes state the closing date; scored at deadline
- Why it matters
- Teams using standalone Claude Design have a fixed date to move their work or workflows elsewhere.
- Try this
- List any design work stored only in claude.ai/design and plan to export it before December 14, 2026.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
VerifiedProvisionalresearch
Anthropic gave six METR staff access roughly like its own risk teams have. METR will publish its risk assessments; Anthropic cannot veto publication but can require redactions in set categories, and decides what falls in them.
How we checked · Why it matters
- How we checked
- Anthropic's announcement page confirms six METR staff, publication rights and redaction terms
- Why it matters
- Sets a public reference point for how much access independent evaluators of frontier models get.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
Docs-verifiedProvisionaltool
Claude Code v2.1.302 turns fork subagents on by default in claude -p, the Agent SDK, the desktop app and VS Code. Setting CLAUDE_CODE_FORK_SUBAGENT=0 turns them off.
How we checked · Why it matters · Try this
- How we checked
- Read official GitHub release notes for v2.1.302; no sandbox run
- Why it matters
- Headless and SDK pipelines may now spawn fork subagents they did not before, which can change cost and behavior.
- Try this
- If you run claude -p in CI, pin the version or set CLAUDE_CODE_FORK_SUBAGENT=0 until you have compared output and token use.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
Docs-verifiedProvisionaltool
Codex CLI 0.163.0 runs independent read-only tools in parallel, lets you send a follow-up to steer Codex while it is still responding, and adds a searchable /config panel.
How we checked · Why it matters · Try this
- How we checked
- Read official openai/codex GitHub release notes for 0.163.0; no sandbox run
- Why it matters
- Parallel lookups and steering mid-turn cut wait time and wasted turns in long agent sessions.
- Try this
- After upgrading, send a short correction while a long Codex turn is running and see whether it changes course without restarting.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
OpenAI (and Sam Altman's blog)primary source
Docs-verifiedProvisionaltool
Cursor released an Environments API in beta for creating and managing cloud agent environments in code. Environments now hold up to 300 repos.
How we checked · Why it matters · Try this
- How we checked
- Cursor's changelog confirms API beta and 300-repo limit; 5x speed is Cursor's own figure
- Why it matters
- Teams can script agent environments for big multi-repo setups instead of configuring them by hand. Cursor's load-speed figure is its own claim.
- Try this
- Read the Environments API beta docs and list which of your repos one shared agent environment would need.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Cursor Changelogprimary source
Docs-verifiedProvisionaltool
Claude Code Review, which reviews GitHub pull requests on Anthropic's infrastructure, is available on Team and Enterprise plans. It has Quick, Standard and Deep levels, and it never approves or blocks a PR.
How we checked · Why it matters · Try this
- How we checked
- Claude Code docs confirm Team/Enterprise availability and the three review levels; no hands-on run
- Why it matters
- Gives teams a managed reviewer with a cost dial per repo that cannot gate merges by itself.
- Try this
- If an org Owner has enabled Code Review (each review is billed), add a REVIEW.md at the root of one repo's default branch describing what to flag, then comment @claude review (you need write access to the repo). If the repo also runs the Claude Code GitHub Action, that comment may trigger it too.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Latent Space (swyx)primary source
Docs-verifiedProvisionaltool
Perplexity open-sourced SPACE under Apache 2.0. It is a sandbox platform that runs code in isolated Firecracker microVMs, meant for agent execution environments.
How we checked · Why it matters · Try this
- How we checked
- GitHub repo exists, README and Apache 2.0 license confirmed; isolation not tested by us
- Why it matters
- A permissively licensed, self-hostable option for isolating code your agents run.
- Try this
- Read the SPACE README and compare its isolation model with the sandbox your agents use now; do not deploy without your own review.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Latent Space (swyx)primary source
Docs-verifiedProvisionaltool
Claude Code v2.1.301 adds classifier_verdict and permission_mode to OpenTelemetry tool_decision events. This lets you tell auto mode classifier decisions apart from permission rule decisions.
How we checked · Why it matters · Try this
- How we checked
- Read official GitHub release notes for v2.1.301; no sandbox run
- Why it matters
- Teams auditing agent permissions can now tell whether each tool decision came from the auto mode classifier or a permission rule.
- Try this
- If you export Claude Code telemetry, add a dashboard filter on classifier_verdict to count classifier blocks separately from rule denials.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
Docs-verifiedProvisionaltool
Google added a Biomolecular Structure Viewer to Antigravity. It is a shared 3D view: the agent can highlight regions of a protein structure, and the researcher's highlights and pinned notes sync back to the agent.
How we checked · Why it matters
- How we checked
- Google DeepMind blog post confirms the viewer; no hands-on test
- Why it matters
- An example of agent tools with a shared visual state that syncs between the person and the agent.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Google DeepMind / Google AIprimary source
Docs-verifiedProvisionaltool
Simon Willison built Chirp Pong with Claude Code and Claude Opus 5.5 Extra. It is a static web pong game for two phones that sends game state as audio chirps between the phones' speakers and microphones.
How we checked · Why it matters · Try this
- How we checked
- Tool page exists and describes audio-chirp play between two phones; not run
- Why it matters
- Shows networking over sound in a static page: the tool page says no Wi-Fi, Bluetooth or accounts.
- Try this
- Read the published prompt on Simon's post as a template for asking an agent to build a small static web app.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Simon Willisonprimary source
Vendor claimProvisionalresearch
Anthropic says Claude, in a closed loop with an automated evaluator, designed better quantum circuits on eight tasks. The arXiv abstract reports significant reductions on seven of the eight. Anthropic's post says one point-addition design uses 12% more qubits but 17x fewer gates.
How we checked · Why it matters · Try this
- How we checked
- Anthropic post and arXiv paper 2610.22367 exist; abstract does not name Claude; results not independently reproduced; an earlier draft of this summary (2026-10-16-57503e5e) failed the full audit and was corrected, then re-audited
- Why it matters
- Shows a reusable pattern: a model iterating against a strict correctness checker and cost score, with little human input.
- Try this
- For an optimization task you own, write an automated correctness check and cost score first, then let an agent iterate against them.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary sourceprimary source
Vendor claimProvisionalresearch
Anthropic reports that Claude worked out how the NIST P-curve seeds were generated: lrand48 seeded with Unix timestamps from 1997 to 1998, not a hash. Anthropic published code to reproduce them.
How we checked · Why it matters
- How we checked
- Anthropic research page confirms the claim and code; we did not run the reproduction
- Why it matters
- A concrete example of a model cracking a long-open puzzle by recognizing weak generator output. Anthropic says no action is needed for curve users.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
Vendor claimProvisionaltechnique
Anthropic replaced git for its monorepo with Scion, an in-house version control system built around a server that is always online. Merges rose from about 2,500 to over 30,000 per day between May and September 2026.
How we checked · Why it matters · Try this
- How we checked
- Anthropic engineering post confirms the system and figures; internal metrics are not independently checkable
- Why it matters
- An early sign of where git setups strain when agents write most of the code, plus ideas for server-side repo APIs. Anthropic says Scion is not a suitable replacement for git for general use.
- Try this
- Check whether your CI jobs do full clones just to answer history questions; see whether your git host's API can answer them, or use sparse checkouts for jobs that need only some directories.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-10-30
Anthropicprimary source
Today's list · All days · latest.md