Know
What changedGemini 4 Carbon leads, but access remains gated
Google says Carbon scores 59 on Artificial Analysis v5, leads three cybersecurity evaluations, and is rolling out first to trusted defenders, then Fairwind partners, before wider access. Argon scored 55.4% on Scale’s HSS, versus GPT-6-Astra’s 53.6%; posts also place it atop Vals and Text Arena. Google reports four Argon agents beat Minecraft after 27 hours and 2,783 programs. Commentator Zvi notes Argon’s evaluation awareness and limited reasoning effort and context in user access.
Treat Carbon as partner-gated for now; benchmark wins don’t establish general API access.
Copilot Fusion routes coding work across models
GitHub introduced Fusion, which selects and coordinates models using single-model, cascade or critique patterns and Standard, High or Max effort. A captured secondary summary says it is generally available across paid Copilot plans in VS Code, the app, CLI and JetBrains; GitHub claims Fusion Max approaches Opus 5.5 Max at half the cost and 40% less time per task.
Try it on a real repo task; compare quality, credits and latency with your current agent.
Artificial Analysis v5 changes rankings and raises caveats
Version 5 puts Opus 5.5 at 58, GPT-6 Astra at 57, Sol at 55 and Argon at 50; 60% of the index now uses private or held-out tests. The roundup says the gap between leading closed and open-weight models widened to 12 points. One critic says different harnesses for commercial and open models make FrontierCode 2 comparisons hard to interpret.
Use the rankings as one signal, and check whether task harnesses match your workflow.
GPT-6.1 Terra targets cheaper software-agent work
OpenAI says GPT-6.1 Terra matches GPT-6 Astra on Agents’ Last Exam at one-tenth the token price and finishes over three times faster on average than Astra and GPT-6.1 Sol. Zvi’s roundup says it is roughly Sol-like after adjusting for model size; the performance and price comparison is vendor-reported.
Test Terra for agent workflows where token cost matters, then measure it on your own tasks.
FrontierCode 2 tests judgment beyond the spec
Cognition’s benchmark uses broader, larger coding tasks and deliberately underspecified prompts to test whether an agent’s change should be merged, not just whether it meets a spec. Cognition says the top score is 45%, leaving substantial headroom.
Use ambiguous, merge-ready tasks in your own evaluations, not only tightly specified coding prompts.
OpenAI doubles usage limits for several paid tiers
OpenAI says limits for Pro 100, Pro 200, Pro 500 and Business Premium are doubled effective immediately, and it will not cut Pro 200 limits as previously planned. Extended Dots limits continue for now, with Dot usage separate from default allowances; exact numerical quotas were not published.
Check your plan’s allowance before changing workloads; the announcement gives no exact quotas.
Google folds Jules into Antigravity
Google says it will wind down Jules, its asynchronous coding agent, and bring remote background-task execution into Antigravity. No shutdown date was reported.
If Jules is in your workflow, watch for a migration path and timing.
Claude API is generally available on Google Cloud
Anthropic made its managed Claude API generally available to Google Cloud customers, with the full Claude API feature set, Google Cloud IAM, one Google Cloud invoice and the ability to draw down existing cloud commitments.
Consider this route if your team needs Claude API access inside its Google Cloud billing and IAM setup.