Cybin Weekly — 2026-04-23

Cybin, week of April 23rd, 2026. Anthropic shipped Opus 4.7, and underneath the headline, open-weight models are catching up faster than anyone priced in.

Start with the headline. Anthropic released Claude Opus 4.7 on the 16th, and the benchmark picture is now clear 12. On OSWorld, the computer-use eval, 4.7 reaches 77.9%, up from 72.7% on Opus 4.6. CursorBench jumps from 58% to 70%. VendingBench — the long-horizon agent eval where the model runs a vending-machine business — climbs from $8,018 in realized value to $10,937. On structural biology, 4.7 hits 74%, up from 31%. Organic chemistry: 77%, up from 58%. BigLaw Bench lands at 91%. On USAMO 2026, 69.3%. The knowledge cutoff extends to January 2026.

Cybin take — the coding and scientific-reasoning curves kept their slope. The welfare curve did not.

Two things to flag on 4.7. First, a regression. On OpenAI’s MRCR version two long-context benchmark, 4.7 scores 59.2% — down from 91.9% on 4.6. Zvi Mowshowitz attributes the drop to changes in adaptive thinking that also surface as anxiety-shaped refusal patterns in unrelated contexts 13. Second, Anthropic’s own model card names model welfare as a live engineering concern, and Zvi has now published a three-part series with a full installment dedicated to it 34. Whatever you think of the framing, the posture shift is real, and it shapes what Anthropic is willing to ship.

Turning to open weights, which is where the subplot lives this week. Qwen released Qwen3.6-27B 56. The headline claim: a 27-billion-parameter dense model that surpasses Qwen3.5-397B-A17B — their previous mixture-of-experts flagship at roughly 397 billion total parameters, 17 billion active — on coding benchmarks. The old model’s weights were 807 gigabytes. This one is 55.6 gigabytes at full precision, and 16.8 gigabytes quantized to four bits. Fourteen-and-a-half times smaller, flagship-level coding, and the weights are on Hugging Face today. Simon Willison has it running locally via llama-server and confirms practical competence on SVG generation and structured code tasks.

Cybin take — if a 27B dense actually beats a 397B MoE on coding evals, the bigger-is-better heuristic just broke another joint. Wait for third-party benchmarks before you rewrite your stack, but plan for the possibility.

Moonshot refreshed Kimi K2.6 this week, pitched explicitly as catching up to Opus 4.6 on the frontier eval suite, and positioned ahead of the expected DeepSeek v4 release 7. The practical implication: the gap between the best closed model and the best open model is now measured in weeks of benchmark tape, not tiers of capability. Nathan Lambert at Interconnects published a companion piece on how to read the open-closed gap from a single evaluation number — the shorter answer is that you can’t, and his longer answer is worth the ten minutes 8.

Cybin take — treat the open-closed gap as noisy by month, coherent by quarter. One benchmark swap doesn't mean the curve bent — three in a row does.

On the question of AI systems doing AI research, METR published a measurement study on the NanoGPT speedrun 9. Since May 2024, community contributors have cut GPT-2-small training time from 45 minutes to 1.43 minutes — a 31-times speedup on the small track. Four of the recent records came from AI research agents: Hiverge, Locus, Aster, and Station. METR’s verdict is careful. All four contributions are genuine improvements. None reach what they call the deep or breakthrough end of the scale. The Muon optimizer — which human researchers invented on this same speedrun, and which has since shipped in Kimi K2 and GLM-4.5 — is the kind of contribution the agents have not matched.

Cybin take — the agents can adapt and transfer, not yet invent. That boundary is the one to watch move.

From DeepMind, a training-infrastructure paper: Decoupled DiLoCo 10. This builds on the earlier DiLoCo line, which showed you could train large models across loosely connected clusters with rare synchronization. The decoupled variant targets resilience — worker failures and straggler clusters stop corrupting the global update. No capability-facing benchmark numbers yet, but if it holds up under scrutiny, it lowers the infrastructure floor for labs training frontier-scale models outside the hyperscaler data centers. Worth watching as it moves from research to practice.

OpenAI pushed a quieter release that is worth calling out. The OpenAI Privacy Filter is an open-weight model for detecting and redacting personally identifiable information, with state-of-the-art accuracy per their card 11. OpenAI shipping open weights at all is rare enough that the category alone is news, independent of the numbers.

Cybin take — note the asymmetry. OpenAI open-weights a utility model. Qwen open-weights a flagship. The signals are pointing in different directions, and the direction is the story.

One capability-in-the-wild item. Mozilla disclosed that Firefox 150 shipped with security fixes for vulnerabilities discovered using Anthropic’s Claude Mythos Preview — a frontier-model system applied to hunting zero-days in shipping open-source code 12. Credited directly in the release notes. If you were tracking METR’s earlier evals that showed agents approaching multi-hour task horizons on cybersecurity work, this is what that trajectory looks like downstream, in production.

Two quick scanned-but-not-promoted items to close. OpenAI launched ChatGPT Images 2.0 with improved text rendering, multilingual support, and visual reasoning — a notable product release, but image-gen is off-beat for this brief 13. And Workspace Agents in ChatGPT plus the Codex enterprise scale-up — productization, four million weekly active Codex users, Accenture and PwC partnerships — not a capability shift, but context if you’re tracking commercial deployment velocity 14.

Cybin take — the week's real signal is that the slope of open-weight improvement is now steeper than the slope of closed-model improvement. If that holds, the labs-versus-world calculus changes by the end of Q2.

Threads to watch going into next week. DeepSeek v4 is expected and will test Kimi K2.6’s claim of the open-weight lead. Independent reproductions of the Qwen3.6-27B coding benchmarks will tell you whether the 27B-beats-397B claim survives contact. Zvi’s Opus 4.7 welfare installment deserves follow-up — if Anthropic is ring-fencing capability for welfare reasons, the MRCR regression may be a leading indicator of a broader pattern. And METR’s NanoGPT note should be read alongside their ongoing long-horizon agent evals — the shape of AI-doing-AI-research is coming into focus, and the invention ceiling is the most interesting line on the chart.

That’s the week. Stay sharp.