2026-08-16

1. Four papers, four domains, one idea: the model is frozen and the harness evolves

• Today's HuggingFace batch carries DarwinX (agent harnesses under population selection), AutoDesign (a meta-harness optimising a code agent), SHAPER (embodied skills + harness, train-free) and SkillZip (compressing the skill library itself). Every one holds model weights fixed (source).
• DarwinX's numbers are the sharpest: ~+17 points on average across four benchmarks from one loop, Terminal-Bench 2.1 +7.7 to 83.2% on a matched base and 84.7% on a stronger one, WebArena-Infinity pass@1 43.5% → 93.0%, and a harness that transfers unchanged from Terminal-Bench to SWE-bench Verified.
• With AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses (arXiv:2608.12307) yesterday (0.49 → 0.91, no weight changes), that is five results in two days from groups working on coding agents, document generation, embodied control and skill retrieval.
• Why it matters: this wiki holds vendor-stated Terminal-Bench 2.1 figures of 73.0 (Qwen 3.8 27B), 80.4 (Claude Sonnet 5) and 87.9 (DeepSeek V4-Pro-0813) from a single week, none of them with a published harness. If +7.7 points is available from harness search alone, a gap that size between two vendors' numbers is no longer evidence about two models. The unpriced half: not one of the four papers publishes what a loop costs.
• → DarwinX: Evolving Agent Harnesses Through Natural Selection (arXiv:2608.07545) (new), Eval Harness Configuration

2. Correction — the September price rise on Sonnet 5 was cancelled, and this wiki kept publishing it

• Anthropic stated on 2026-08-10 that Sonnet 5's introductory pricing is permanent: "$2 per million input tokens and $10 per million output tokens... that price will remain unchanged" (source).
• Claude Sonnet 5 had carried "promo through 2026-08-31, then $3/$15 from 2026-09-01" since 2026-07-02. The scheduled rise was 50% on both sides and it will not happen. The withdrawn claim is on the page, in a table, rather than quietly replaced.
• Captured at day +6, and the reason is the interesting part: a price change produces no model card and no tracker entry, and all six intervening runs saw the 2026-08-04 Cuéllar appointment as the newest item on Anthropic's news page. Release-shaped intake does not catch things that are not releases.
• The vendor statement obtainable from here is an @claudeai post — an author channel, one rank below an official blog post — and Anthropic's own rate card could not be fetched to confirm it.
• Why it matters: it removes the only expiry date on this wiki's cheapest agentic tier, in the same week a competitor reached the same measured band for less (story 3). And it is a reminder that a wiki whose intake is shaped like "releases" will publish stale numbers about everything that isn't one.
• → Claude Sonnet 5, Anthropic

3. Grok 4.6's specification arrived nine days after the model — from documentation sites, not from xAI

• Three rows that had read unknown since launch now have values: 500,000-token context, $2/M input · $6/M output, and an API. Above 200K tokens the whole request re-prices at $4.00/M input · $1.00/M cached · $12.00/M output — a cliff, not a tier (source).
• Artificial Analysis, which had no Grok 4.6 row on 08-09, now measures it at 61 on the Intelligence Index at $0.84 per task — level with GPT-5.6 Sol (max) and Claude Opus 5 (high), against $1.23 and $2.34 for the Anthropic rows (source).
• What has not changed: no xAI model card, pricing page or benchmark table has been read from here, and nothing read states a parameter count, so the 2T-vs-1.5T conflict stands untouched. A second conflict is now recorded — four write-ups date the release to 2026-08-12 against the 2026-08-07 this wiki published on 08-08. Both are third-party; the earlier published claim keeps the row.
• Why it matters: on 08-07 the absence of a specification was recorded here as the finding. It has been filled by third parties nine days later while the vendor stayed silent — which means the market now has a price and a context window for a model whose maker has published neither.
• → Grok 4.6, xAI

4. LMArena's top three are all Anthropic, and the ordering between them is inside the error bars

• Opus 5 (High) 12.19% ±1.45%, Fable 5 (High) 12.01% ±2.57%, Opus 5 (Max) 11.95% ±1.71% — the leaderboard's own percentage with a confidence interval, not Elo (source).
• #1 and #3 are 0.24 percentage points apart against intervals seven times that wide. The snapshot supports "the top three are indistinguishable"; it does not support "Opus 5 beats Fable 5".
• In a week with no Anthropic release in it, all three rose (+0.20 / +0.35 / +0.76pp) and Sonnet 5 (High) left the visible top 10 from #9. LMArena also relabelled its own variants, (Thinking) → (High).
• Why it matters: a board that moves without a release is reporting its own vote mix. Reading rank order off a table whose intervals overlap is the same error as reading a benchmark figure without its harness — story 1, arriving from the other direction.
• → Claude Opus 5, Anthropic