Weekly Synthesis — W33 (2026-08-10 → 2026-08-16)

Period

The week capability stopped arriving in the weights. Four of the week's releases were explicitly the same model, post-trained harder — and on Sunday four unrelated research groups published the strongest version of the same claim: leave the weights alone entirely and evolve the harness. A field that spends its announcements on parameter counts spent this week demonstrating that the parameter count was not the variable.

Notable Releases

• Gemini 3.7 Flash (2026-08-13) — GA 23 days after 3.6 Flash with every capacity figure identical: 1M context, 64K output, March 2026 cutoff. What moved was the benchmark column — AutomationBench 17.0% → 30.4%, DeepSWE 48.6% → 65.3%. Logan Kilpatrick attributes it to algorithmic improvement over roughly three weeks rather than scale. Priced $0.75/M · $3.75/M introductory through 2026-12-31, reverting to $1.50/M · $7.50/M.
• DeepSeek V4-Pro-0813 (2026-08-13) — V4-Pro leaves preview as a dated build. Vendor-stated Terminal Bench 2.1 72.1 → 87.9 and DeepSWE 12.8 → 62.7, with no published harness; the one independently produced figure, Artificial Analysis' index, moves one point to 53.
• GLM-5.3 (2026-08-14) — Z.ai states it reuses GLM-5.2's base exactly as it was, every gain from post-training. Terminal-Bench 3.0 4.6 → 28.3, CyberGym 84.5%, Agents' Last Exam CLI 28.5 against GPT-5.6 Sol's 28.6. Marketed as "the strongest open-weights coding model" with the weights stated for roughly two weeks out, behind a safety evaluation.
• Qwen 3.8 27B (2026-08-14) — shipped four days after its announced date, 27.78B dense, Apache 2.0, 262,144-token context. Its SWE-MM row is run on the Claude Code harness, cited to Appendix 8.3 of the Claude Opus 4.7 system card — a Chinese lab naming a competitor's harness by document number, and more harness detail than any other release this week published for any figure.
• Qwen 3.8 Max open weights (2026-08-12) — the first Max-class Qwen opened, as Qwen3.8-2.4T-A95B, 24 days after the closed preview.
• Muse Glimmer (2026-08-10) — Meta released the student under Apache 2.0 and kept the teacher closed.
• GPT-5.6-Cyber (2026-08-10) — a cybersecurity model trained to refuse less on dual-use work, at Daybreak Red tier only, no published price and no system card.
• Grok 4.6 — no release of its own this week, but its specification finally became readable: 500K context, $2/$6 per M, and an Artificial Analysis index of 61 at $0.84 per task, all of it from third parties nine days after launch.

Emerging Themes

1. The harness, not the weights

The week's clearest signal is a convergence nobody coordinated. On 2026-08-15 AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses (arXiv:2608.12307) reported a builder-written inference harness moving a weaker model's Theory-of-Mind average 0.49 → 0.91 with no parameter updates. On 2026-08-16 four more arrived at once — DarwinX: Evolving Agent Harnesses Through Natural Selection (arXiv:2608.07545) (harness populations under selection, ~+17 points average, Terminal-Bench 2.1 to 84.7%), AutoDesign (meta-harness optimisation, PosterBench 54.99 → 67.39), SHAPER (embodied, train-free) and SkillZip (+12.2 points at 3.46× skill-library compression). Coding agents, document generation, embodied control, retrieval: four domains, one structure, model frozen throughout.

Set that beside the releases. GLM-5.3 is GLM-5.2's base with more post-training. Gemini 3.7 Flash is 3.6 Flash's capacity with a better benchmark column. DeepSeek shipped a dated build, not a new model. Four vendors and four research groups said the same thing in one week from opposite directions: the number moved and the model did not.

The consequence lands on Eval Harness Configuration. This wiki holds three vendor-stated Terminal-Bench 2.1 figures from a single week — 73.0, 80.4, 87.9 — with no published harness between them, and DarwinX reports +7.7 points on a matched base from harness search alone. A spread that size is no longer evidence about two models.

2. "Open weights" became a date

Open-Weights Policy Fight gained its first entry that is a promise: GLM-5.3 claims the open-weights crown with nothing downloadable, weights due back around 2026-08-28. Qwen ran the completed version of the same sequence — closed preview, paid API, open weights, 24 days — and slipped its second checkpoint four days without ever restating the date in a first-party channel.

The week also produced the measurement that had been missing. Artificial Analysis now lists both forms of Qwen3.8 and scores them identically: Max 58, open 2.4T A95B 58. Whatever the schedule is for, it is not a quality discount — the artefact handed over is the artefact that was sold. What is bought by the delay is a paid window, and nothing published this week distinguishes that reading from "a safety evaluation takes two weeks".

3. The missing denominator

A pattern visible only across the week: almost every headline figure shipped without the second number that would let a reader judge it. Anthropic's auto mode published 89% recall and no false-positive rate (though it did publish the sharper paired figure — 800 commands blocked that humans approved against 6 the reverse). SL2T shipped with no accuracy figure. OpenAI's Ultrafast claims "the same intelligence as Sol Standard" with no benchmark and no price. GLM-5.3 has no per-token price. Grok 4.6 went nine days with no model card. Conceptual Reasoning Index (CRI) scores reasoning about unverifiable questions without saying how such a benchmark is itself validated. Mechanist publishes no number at all. DarwinX publishes no compute cost for the loop that produces its +17 points.

Declining Themes

• Parameter count as the headline. Not one release this week led with one, and two of them — GLM-5.3 and Gemini 3.7 Flash — went out of their way to say nothing had been trained. The 2T-vs-1.5T dispute on Grok 4.6 is now nine days old with no vendor page to settle it, and nothing published this week mentioned it.
• The Chinese-labs-are-behind framing. It did not survive the week's own measurements in either direction: Qwen3.8's index of 58 sits below every model this wiki calls frontier, while its open checkpoint matches its paid one exactly. Both facts are duller and more useful than the framing they replace.

Surprising Results

• The opened Qwen checkpoint measures the same as the API it was cut from — 58 against 58, $1.09 against $1.13 per task. The only difference is a 984k context window against a stated 1M, which is a property of whoever serves the weights.
• DeepSeek's own numbers and the independent one disagree by two orders of magnitude of movement — vendor-stated gains of 15 to 50 points against an Artificial Analysis move of exactly one. Both are on the page; neither was averaged.
• LMArena's top three are all Anthropic and their ordering is inside the error bars — 12.19% ±1.45%, 12.01% ±2.57%, 11.95% ±1.71%, with #1 and #3 0.24 percentage points apart. All three rose in a week with no Anthropic release in it.
• Anthropic cancelled a scheduled 50% price rise on Claude Sonnet 5, making $2/$10 per M permanent — the first price this year to move downward by not moving at all.

Open Debates

• Is a staged open-weights release a norm arriving, or a paid window with an open label? GLM-5.3 gives a safety evaluation as the reason for a two-week gap; the same lab put subscriber access and open weights three days apart in June. Nothing published separates the readings, and 2026-08-28 is when the question gets an answer rather than an argument.
• Does a benchmark figure without its harness mean anything? The week supplies evidence on both sides: Qwen disclosed a harness for one row of four, and four papers demonstrated that the undisclosed term is worth more than the gaps between vendors' published numbers.
• Who checks a finding an automated researcher produced? Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence (arXiv:2608.12036) reports a mechanism theory of belief and a cross-modal safety risk, discovered end-to-end by an agent, with no independent validation described — the question Automated Weak-to-Strong Researcher (AAR) raised, now attached to concrete claims.

Outlook

Two dated things to check. GLM-5.3's weights are due back around 2026-08-28, which turns Open-Weights Policy Fight's newest entry from a promise into either an artefact or a pattern. And Grok 4.7 is targeted at roughly 2026-08-22 on xAI's stated monthly cadence — with 4.6 still lacking a model card, the thing to watch is not the date but whether a specification ships with it.

The larger question the week opens is what a model page is for. This wiki's ## Spec and ## Benchmarks tables assume the model is the unit of capability. Four papers and four releases just argued it is not — and neither the harness nor the router that now carry the difference has a page, a version, or a published error rate anywhere in the field.

Related

• Weekly Synthesis — W32 (2026-08-03 → 2026-08-09)
• Eval Harness Configuration
• Open-Weights Policy Fight
• Model Routing
• DarwinX: Evolving Agent Harnesses Through Natural Selection (arXiv:2608.07545)