~ / news

AI Coding Agent News

Recent market notes for ranking changes, pricing shifts, model releases, and benchmark risk, refreshed 2 November 2026. Rumors and source quality are clearly marked.

Nov 1, 2026Google ships Gemini 4 Carbon: an Artificial Analysis v5 score of 59 leads every independent measurementThe mid-tier of the staged Fairwind release lands above Claude Opus 5.5 (58) and GPT-6 Astra (57) on v5, the first top-of-table figure that comes from an independent index rather than a maker sheet. Google’s own release claims ExploitGym 54.8, Collinear CWE-bench 71.0 and first place on 15 of 17 tests; none of that is independently reproduced yet, so the board lists the model as not yet tested on DeepSWE. The leaderboard row.TypeModel releaseSourceOfficial + Independent
Oct 27, 2026Artificial Analysis v5 puts 60% of scoring behind private held-answer tests: Xiaomi falls 16.3 points, GLM-5.3 13, Kimi 9The largest drop landed on the model marketed as the strongest open source at 46.32 on v4.3, now 30.0 on v5. The two best released open-weight models sit at 32, below Claude Haiku 5.5, and the open-weight gap to the leader widened from 3 to 12 points. Artificial Analysis cautions that the redesigned benchmark, not a proven general fall, drives the widened gap. The full drop analysis.TypeBenchmarkSourceIndependent
Oct 18, 2026A 50% targeted-training penalty lands on DeepSeek V4.1 Flash: its cost coefficient drops 1.10 to 0.55The audit: 74.2 is the vendor model-card figure under mini-SWE-agent, while the same model runs 65.6 under Codex and posts 30.0 on Terminal-Bench 3.0 against Opus 5's 43.3. A number that survives only in the maker's own harness is a demo, not a measurement. The raw 74.2 stays visible on every source table; the dagger halves its weight in the composite and cost model.TypeBenchmarkSourceIndependent
Sep 21, 2026AgentRanks moves the board to DeepSWE 1.1: Muse Spark 1.3 leads at 75.4, Claude Opus 5 re-measures at 68.8The 1.1 refresh re-measured every model on harder long-horizon tasks and reordered the top ten: GPT-6 Astra second at 74.1, Gemini 3.8 Flash third at 73.7, Opus 5, the 1.0 leader at 73.6, ninth at 68.8. DeepSeek V4.1 Flash entered at 74.2 that week; an 18 October audit traced that figure to the vendor model card and penalized it (see the two items above).TypeBenchmarkSourceIndependent
For traffic and trust, AgentRanks should treat news as ranked impact notes: what changed, which agents/models it affects, and whether the source is official, third-party, directory-listed, or speculative.