Codex Astra went 18-0 in StarCraft — and still played like a beginner
A new round-robin benchmark put 19 model configurations through 171 games of real-time StarCraft: Brood War. The best of them never lost — and its own author says none of them played past beginner level. The reason is the most useful thing in the results.
Codex Astra won all 18 of its matches in Brood War Bench, a new agent benchmark that has 19 model-and-effort configurations play each other in real time, and its author's conclusion is blunt: no model in the field played past a beginner level. The run is a full round robin — 171 matches, each model driven through its own vendor CLI (Claude Code, Codex, Grok Build) over a minimal observation-and-command API, with matches running in parallel on Freestyle VMs. The top line is Codex Astra at high effort: 18 wins, 0 losses, 12.6 actions per minute, $10.54 a game. Claude Fable took third at 15–3 and $12.24, Codex 5.6 Sol at medium took fifth at 13–5 for $5.12, and the bottom of the table is brutal — Claude Haiku went 0–16 at 0.3 actions per minute, and two Grok 4.6 configurations won one game between them.
The interesting finding is not who won. It is that thinking time is the bottleneck in a real-time environment, and the models that reason least sometimes play best. Older models treated Brood War as turn-based: they issued an order, then reasoned while their base was demolished. Grok 4.6 at high effort logged 11,138 reasoning tokens in a single 43-minute game, issued six command batches, and never fielded a combat unit. Newer models were more aware of the cost of thinking, but not immune — which the author offers as the likeliest explanation for lower-effort settings sometimes beating higher ones. Codex 5.6 Terra at low effort averaged 48.3 actions per minute for $4.65 a game; the same model at high effort averaged 15.8 for $2.10. Effort bought reasoning depth and lost the match.
The second failure is coordination, and it is the one that should worry anyone running multi-agent systems. Codex frequently spun up separate subagents for economy, army production and army control, and they barely talked to each other — so the army agent fed units into the enemy one at a time, unaware of the larger force the other agents thought they were building. That is the classic beginner mistake, and it is exactly the failure mode that shows up when you decompose a job across agents without a shared plan. The author notes that when he directed Codex himself, it planned attack timings far better.
The caveats are real and the author does not hide them. It is a single-author benchmark, and harness and model are entangled by design — each model ran inside its own vendor's agent CLI, so a difference in results is not cleanly attributable to the model. There is one match per pairing, so the 100% win rate rests on 18 games, not hundreds. His own summary of the ceiling: a human beginner playing a photon rush would win every one of these games.
A new cryptography paper closes the gap that made image-provenance proofs useless in practice — lossy compression. Camera attestation signs a photo at capture, but every normal edit breaks that signature: JPEG encoding, blurring a face, redacting a document. ZK-JPEG, from Samuel Dittmer, Steve Lu, Kimberlee Model and Joseph Near on the IACR ePrint archive, proves that a published JPEG was correctly compressed from a secret, committed input, and folds a family of image transformations into the same proof at minimal extra cost, using the PicoZK toolchain over the line-point zero-knowledge proof system. Earlier zero-knowledge provenance work proved an edit history but did not survive JPEG, which is the format most photos actually travel in.
The redaction case is the one that matters. Today's provenance push — signed capture metadata and machine-readable marks that travel with a file — assumes the honest publisher ships the original. ZK-JPEG's framing is that you can blur or cut a region to protect the subject and still prove the surviving pixels came from the committed source, instead of choosing between privacy and provenance. It is a proof system, not a product: no implementation, no deployment, and the cost of proving a full-resolution photo is not settled by the abstract. If you want the background on why this problem exists at all, AI 101 — What is a deepfake? covers the detection-versus-provenance split, and What is AI watermarking? goes through the marking side.
What to watch: whether a provenance standard picks up a compression-aware proof, and whether anyone re-runs Brood War Bench with one shared harness to separate model skill from agent scaffolding.
If reasoning costs you the game in a real-time environment, what does that say about the agents you run on deadlines? Tell us in the comments.
