OpenAI changed two API settings and tripled its model's score on a benchmark it had been failing. A cluster of Claude Code changelog entries mostly concerned what agents may do to a machine. A harness vendor's near-perfect result on the same benchmark turns out never to have been independently verified. If you cannot say which harness produced a number, you do not know what the number measures.
1. Two API settings tripled a frontier model's score on a benchmark it was failing
How enabling two settings tripled our scores on the ARC-AGI-3 benchmark · Ilan Bigio and Ted Sanders, OpenAI · openai.com, 29 July 2026
ARC-AGI-3 asks an agent to work out how an unfamiliar 2D game works without being told. GPT-5.6 Sol scored 7.8% on it, and OpenAI went looking for why. The answer was not the model. ARC's harness discarded the model's private reasoning after every action, so it re-derived the game from scratch each turn, and it handled context limits with a rolling truncation window at 175,000 characters, so it also lost its record of past moves. Reimplemented on OpenAI's Responses API with reasoning retained and compaction enabled, the same model went from 13.3% to 38.3% on the public task set, with roughly six times fewer output tokens. OpenAI estimates the average human tester at 48%. Bigio and Sanders: "Benchmarks rarely measure AI models in isolation. They also measure less visible choices about API settings, harness design, and prompting." ARC's generic harness is a deliberate design choice, on the reasoning that it makes model shortcomings visible and comparisons fair. The three recommendations are specific: use the Responses API rather than Chat Completions, retain reasoning, use compaction.
Why it matters: This is the vendor-side confirmation of what Thomas Schranz measured from the outside this morning, and it is stronger, because it isolates two named settings and moves both score and token count in opposite directions at once. Two consequences. Check whether your own harness drops reasoning between tool calls, because that is a silent, expensive default. And treat any benchmark number without a harness description as uninterpretable. OpenAI has an obvious interest in a finding that its model was underrated and its own API is the fix, and every figure here is its own.
2. Six Claude Code changelog entries, most of them about what an agent may do to your machine
Claude Code CHANGELOG, versions 2.1.215 to 2.1.220 · Anthropic · read in full 20:22 EDT, 1 August 2026
The changelog section spanning versions 2.1.215 to 2.1.220 focused largely on containment. The costliest single line is in 2.1.219: Claude Opus 5 became the default Opus model, with a 1M context window and fast mode at $10 and $50 per million tokens. Two new settings appeared: sandbox.network.strictAllowlist to refuse non-allowlisted hosts without prompting, and sandbox.filesystem.disabled to drop filesystem isolation while keeping network egress control. Three entries addressed agents escaping their own boundaries: worktree-isolated subagents redirecting git into the shared checkout via git -C or GIT_DIR, background sessions not canonicalizing symlinked working directories "which could let sessions escape their workspace folder," and scheduled-task writes following a symlink at .claude to land outside the project. The entries also described a cap of 20 concurrent subagents and a fix for --max-budget-usd not stopping background subagents. Version 2.1.217 stopped subagents spawning nested subagents by default, while 2.1.219 set the default depth to 3.
Correction, August 3: our July 30 edition had already recorded version 2.1.220 with a July 25 date, so these were not six releases in roughly 36 hours. This item is a feature summary of entries observed in Anthropic's mutable documentation changelog, not a reliable release timeline. The GitHub changelog has since been rewritten and no longer preserves the same sequence.
Why it matters: Two things to act on. The observed changelog said --max-budget-usd had not bounded background subagents before the fix. And if you did not pin CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH, the documented fan-out default changed across two entries. The broader lesson survives the chronology correction: write your limits down rather than silently inheriting mutable defaults.
3. A harness claimed 99% on ARC-AGI-3. The verified board's best number is 30.2%
ARC Prize leaderboard · ARC Prize · read 20:25 EDT, 1 August 2026
On 16 July an anonymous team published "Frontier Models with Our Harness Achieve ~99% on ARC-AGI-3 Public," reporting 98.98% against a 42.83% Claude Code baseline on the same models. The Dark Factory sweep read that essay end to end on 17 July and parked the claim as unverified. It closed the question this morning, and the answer is negative: the figures are self-reported public-set results computed from the authors' own run artifacts, and ARC Prize has verified none of them. Read tonight, the leaderboard's highest ARC-AGI-3 entry is Claude Opus 5 at High reasoning, 30.2%, dated 24 July, at $1.45 per task. Every other entry sits at 7.8% or below, most under 1%. Opus 5's 30.2% is roughly four times the previous best and the first double-digit figure on that column.
Why it matters: Item 1 explains exactly how a self-reported harness result gets to be enormous and meaningless at the same time, and this is what that looks like in the wild. The rule to carry: a harness benchmark from the harness's author, on the public set, with no independent run, is a claim about the author's harness and not about the models. Note also that the verified frontier moved without anyone announcing it. Opus 5's 30.2% has sat on that board since 24 July.
Assembled from the Dark Factory landscape sweep of 04:55 EDT, the rising-conversations monitor passes at 12:00, 15:00 and 18:00 EDT, and a live release-watch pass at 20:10 to 20:30 EDT, for the window 1 August 10:15 EDT to 20:30 EDT. Five things limit this edition. Items 1 and 3 rest on numbers published by the parties they flatter or indict, OpenAI's own and ARC Prize's own, with no third-party replication of either. OpenAI reports 7.8% and 13.3% for the same model without labelling which set each belongs to; the 13.3% is stated as the public set and the 7.8% is not stated, so the two are reported here as printed rather than reconciled. The Schema Harness essay in item 3 was read in full by this pipeline on 17 July, not by this edition. Item 2's changelog carries no release dates. And the Cursor cost-visibility reversal reported in this morning's edition, which that edition could not verify, was confirmed here by a direct read of the forum thread at 20:14 EDT: the staff post is marked solved and accepted, the thread stands at 8.8k views and 49 posts. It ran this morning and is not repeated. No thread-watch lines this cycle.