Google releases Gemini 4 Neon at $3/$15, scoring 61 on AA index Google released Gemini 4 Neon at $3 per million input tokens and $15 per million output tokens, scoring 61 on the Artificial Analysis Coding Agent Index, according to Artificial Analysis. The index score puts Gemini 4 Neon in line with Claude Code on Fable 5.2 at roughly half the cost, and its $6.44 per task is about 17% below Gemini 4 Carbon. Coding agent teams get comparable results for about half the spend. Artificial Analysis puts its Coding Agent Index score in line with Claude Code on Fable 5.2 at roughly half the cost, and its $6.44 per task is about 17% below Gemini 4 Carbon. Coding agent teams get comparable results for about half the spend. Read: Artificial Analysis puts its Coding Agent Index score in line with Claude Code on Fable 5.2 at roughly half the cost, and its $6.44 per task is about 17% below Gemini 4 Carbon. Coding agent teams get comparable results for about half the spend. Read: Agent jobs that can wait, such as overnight reports, now get up to 40% off GPT-6.2 Astra and Sol inside the Responses API, with no separate batch pipeline to build. Reusing a continuation ID also moves follow-up requests up the queue. Try: A week after six fixes to the same servers, mcp-server-git still let a repository's own config run commands with the server's privileges during a merge or read. Agents that open untrusted repositories through it carried that risk until this release. Read: Slower inference explains only part of why Carbon runs take longer than Opus and Astra runs. Most of the gap is extra tool calls and re-sent context, about 16x the input tokens, so Carbon's per-token price understates what each fixed bug costs. Read: Human review found real defects in about half of the flagged patches, so a passing test suite overstates how often an agent's fix is correct. BenchProofer builds formal specs and Lean proofs from existing tests, a stricter check teams can run on agent patches. Read: Human auditors scored AUROC 0.59, near chance, and the best frontier model reached 0.73 and fully identified 32% of cases. Handing research to agents and reviewing their transcripts afterward rests on a check that misses most sabotage. Read: Microsoft AI released seven models across code, cyber, image, voice and reasoning. Agents pay per token at every step, so a coding model at 40% of Sol's price cuts that bill by more than half, if outside tests confirm Microsoft's own quality figures.