Try
Useful tools & techniquesOpus 5.5 fast mode opens to API users
Claude Platform fast mode is generally available: set speed to “fast” for the same Opus 5.5, up to 2.5× output speed, at $8/$40 per million input/output tokens. It is also on Microsoft Foundry; Google Cloud is coming. Claude Code fast mode is unchanged. Anthropic’s Ado reports a standout coding day; Theo’s Haiku-dominance claim is opinion. Joe Weisenthal says Opus recognized his novel.
Try fast mode for streamed, output-heavy calls; handle rate limits and compare the premium cost.
Claude Managed Agents adds dynamic workflows
Claude Managed Agents dynamic workflows are in public beta. A lead agent creates a plan that runs across multiple agents in phases, then combines their results; the capture gives no pricing or limits.
Try it for work that naturally breaks into coordinated, multi-step agent tasks.
Know
What changedOpenAI publishes hundreds of AI-generated math results
OpenAI’s internal model generated 722 manuscripts across 372 families of problems; the account says 719 remained after three withdrawals. It reports results on 90 of Proof Atlas’s top 500 open problems, including a matrix-multiplication exponent bound of 2.25. Lean proofs covered 42% of the papers, leaving much of the release less formally checked.
Treat the theoretical results as research claims, not practical algorithm improvements.
Ultramarine 1 Flash targets low-cost knowledge work
Paper Instruments says Ultramarine 1 Flash is available through its API, with weights expected in two weeks. On its internal professional-work benchmark, it claims $0.13 per task, versus $8.19 for Claude Fable 5 Max and $2.09 for GPT-5.6 Sol Max, with scores above both; the benchmark measures finished office work, not coding.
Consider it for document-heavy workflows, but test its benchmark claims against your own tasks.
AI helped build a Chrome full-chain exploit
XBOW says its researchers used public GPT-5.6 Sol to build a full-chain Chrome exploit affecting an estimated billion Windows users, earning Google’s $250,000 bounty. After one bug in an earlier chain was already known, the replacement took six days and $40,000 in compute; the bugs are now fixed. XBOW advises restarting Chrome when its update prompt appears.
Keep Chrome current and assess your own software for exploitable chains, not just individual bugs.
Meta expands its Muse models and coding tools
Meta says Muse Wonder is available through its Model API and Muse Code; Wonder Full is for Muse Code subscribers, with long-horizon coding and memory features claimed. A post quoting Meta announced open-weight Muse Spark at 816B total/42B active parameters and a 1M-token context. Deedy reports Muse retention above 30% at day seven and 22% at day 14; one user says notifications feel spammy and use is only weekly.
Test Wonder in a real coding workflow; treat vendor capability claims and retention figures as unverified here.
Anthropic offers free scans for open-source projects
Anthropic’s opt-in OSS Scanner offers eligible open-source projects periodic scans by its strongest models at no cost. Core maintainers can apply via a repository pull request; reports are model-generated and unreviewed, so may be wrong. Anthropic says expert testers accepted 85 of 97 high- or critical-severity findings; 11 others were real duplicates, and one was invalid. Its critical-infrastructure defense program is inviting interest for expansion in 2027.
Check eligibility, but route every automated finding through your normal validation and disclosure process.
Gemini 4 Argon rolls out across Google products
Google says Gemini 4 Argon is rolling out globally in the Gemini API, to Pro and Ultra subscribers in the Gemini app, and to Ultra subscribers in Antigravity. The company touts coding persistence and up to one million output tokens when needed. A short report says staff are testing another model, Carbon, but gives no further details.
Check your existing Google account or API access; treat the capability claims as vendor-reported.