Daily drop
Checked claims for October 27, 2026
Top claim
VerifiedProvisionalrelease
OpenAI released GPT-6.1 Terra, API id gpt-6.1-terra, priced at $1 per million input tokens, $0.05 cached input and $5 output, with a context window of about 1.05M tokens.
- How we checked
- OpenAI API model docs confirm model id, pricing and context window; GPT-5.6 Terra docs page checked for comparison
- Why it matters
- A midrange option for computer use, vision and writing at roughly half the price of GPT-5.6 Terra ($2/$12).
- Try this
- Rerun one of your own vision or computer-use evals on gpt-6.1-terra and compare cost per task against your current model.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
OpenAI (and Sam Altman's blog)primary sourceprimary sourceindependent source
VerifiedProvisionalrelease
GPT-6.1 Terra is generally available in GitHub Copilot for Pro, Pro+, Max, Business and Enterprise users. It is billed at provider list pricing under usage-based billing, and the rollout is gradual.
How we checked · Why it matters · Try this
- How we checked
- Read GitHub's own changelog entry (primary for its release): plans, billing and admin policy confirmed
- Why it matters
- Copilot users can try a cheaper agentic coding model than GPT-6.1 Sol. Business and Enterprise admins control access through model policy.
- Try this
- Copilot admins: check the model policy in Copilot settings, since new models are enabled by default unless the global default is off.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
GitHub Changelogprimary source
VerifiedProvisionalresearch
ProgramBench (arXiv 2605.03546, May 2026) asks agents to rebuild 200 programs such as FFmpeg, SQLite and the PHP interpreter from only an executable and its docs. At launch, no tested model fully solved any task.
How we checked · Why it matters
- How we checked
- arXiv abstract confirms 200 tasks, setup and launch results; the newer 18.5% Opus 5.5 score was not found on a primary source
- Why it matters
- A hard end-to-end software-building benchmark. Import AI reports a much higher later Opus 5.5 score, not yet confirmed from a primary source.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Import AI (Jack Clark)primary source
Docs-verifiedProvisionaltool
GitHub Copilot assisted permissions are GA in Copilot CLI, the Copilot app and VS Code. An LLM reviews permission requests for commands, file paths and URLs; requests it judges clearly safe run without asking.
How we checked · Why it matters · Try this
- How we checked
- Changelog and GitHub docs checked for behavior, enablement steps and caveats; not run in a sandbox
- Why it matters
- Fewer approval prompts for agents. GitHub's docs warn the reviewer model can approve dangerous actions and is not a substitute for sandboxing.
- Try this
- In Copilot CLI run /permissions and select assisted, but only inside a sandbox; avoid combining it with --allow-all.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
GitHub Changelogprimary sourceprimary source
Docs-verifiedProvisionaltool
Claude Code v2.1.311 makes the ultrathink keyword run its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API. Typed while Claude is working, it raises the rest of that turn.
How we checked · Why it matters · Try this
- How we checked
- Read official GitHub release notes for v2.1.311; behavior not tested in a sandbox
- Why it matters
- Gives a predictable way to raise reasoning effort per turn without changing session settings.
- Try this
- After updating Claude Code, use ultrathink on one hard turn and compare usage in /usage against a normal turn.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Anthropicprimary source
Docs-verifiedProvisionaltool
On Bedrock, Mantle, Vertex and Foundry, Claude Code v2.1.311 now loads model settings from a signed file on downloads.claude.ai. Setting CLAUDE_CODE_MODEL_CATALOG=off turns this off.
How we checked · Why it matters · Try this
- How we checked
- Read official GitHub release notes for v2.1.311; not tested in a sandbox
- Why it matters
- A new outbound request that locked-down or air-gapped enterprise setups may need to allow or disable.
- Try this
- If you run Claude Code through a cloud provider behind a strict firewall, decide whether to allowlist downloads.claude.ai or set CLAUDE_CODE_MODEL_CATALOG=off.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Anthropicprimary source
Docs-verifiedProvisionaltool
Claude Code v2.1.311 fixes standard-context sessions stopping with "Usage credits required for 1M context" when a request ran over 200K tokens; it now compacts the conversation and continues.
How we checked · Why it matters
- How we checked
- Read official GitHub release notes for v2.1.311; not reproduced
- Why it matters
- Long agent sessions on standard context should no longer halt unexpectedly at the 200K boundary.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Anthropicprimary source
Docs-verifiedProvisionaltool
webframe (tractordev) is an MIT-licensed Go CLI that runs a web page in a hidden system webview, sending console.log to stdout and exceptions to stderr, without bundling a browser engine.
How we checked · Why it matters · Try this
- How we checked
- Repo README confirms purpose, Go, MIT license, system webview engines; not run in a sandbox
- Why it matters
- A lightweight option for scripted page JavaScript and scraping without Playwright, though Simon's evaluation found real rough edges.
- Try this
- Read Simon's evaluation and the repo README to see whether the listed rough edges block your use case before trying it.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Simon Willisonprimary sourceindependent source
Docs-verifiedProvisionaltool
Gemini CLI nightly v0.67.0-nightly.20261027 includes MCP server and tool metadata in ACP permission requests and classifies project spending cap errors as terminal quota errors.
How we checked · Why it matters
- How we checked
- Read official GitHub pre-release notes; nightly build, not tested
- Why it matters
- ACP clients can show which MCP server and tool is asking for permission, and runs stop cleanly on spending caps.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2026-11-10
Google DeepMind / Google AIprimary source
Seen but not verified (don't act on these)
Not verifiedProvisionalprediction
Theo predicts that token spend of about $100k per month today will cost around $100 per month within a year for equivalent model capability.
How we checked · Why it matters
- How we checked
- Dated prediction from video notes (Gemini-analyzed); logged for scoring at deadline
- Why it matters
- A falsifiable cost-curve forecast that affects build-versus-wait decisions for token-heavy agent products.
- Source record
- Tier T1-seed · trust 0.5 (lower bound 0.196) · re-check 2027-11-10
Theo (t3.gg)
Not verifiedProvisionalprediction
Jack Clark predicts ProgramBench will be saturated, meaning a very high resolved score, by the end of 2027.
How we checked · Why it matters
- How we checked
- Prediction logged from newsletter; benchmark existence checked on arXiv
- Why it matters
- A dated forecast on autonomous end-to-end software building that can be scored against public results.
- Source record
- Tier T3 · trust 0.5 (lower bound 0.196) · re-check 2028-01-14
Import AI (Jack Clark)primary source
Today's list · All days · latest.md