Cybin Weekly — 2026-04-26

Cybin, week of April 26th, 2026. Two flagship models dropped on consecutive days — GPT-5.5 from OpenAI and DeepSeek V4 from China — and Anthropic published a public postmortem confessing why Claude Code felt off for two months.

Start with GPT-5.5. OpenAI shipped it on April 23rd, alongside a system card and a Bio Bug Bounty offering up to twenty-five thousand dollars for universal jailbreaks targeting biosecurity risks 12. The architectural headline, per OpenAI engineer Romain Huet, is that 5.5 completes the unification of Codex and the main model line that began in 5.4 — there is no separate coding model anymore, and the strongest gains land on agentic coding, computer use, and any task where the model is driving a computer 3. Simon Willison had preview access and confirms the model is fast, capable, and lands a noticeably better pelican-on-a-bicycle than 5.4 in extended-reasoning mode, while losing in standard mode 45.

Now the part that should set off alarms in your budgeting spreadsheet. GPT-5.5 is twice the price of 5.4 — five dollars per million input tokens and thirty per million output, where 5.4 was two-fifty in and fifteen out. The Pro tier runs thirty in and one hundred eighty out 4. The API is not yet generally available; access is currently through ChatGPT subscriptions and the Codex CLI.

Cybin take — capability per dollar is not moving as fast as capability per generation. The frontier is getting smarter and more expensive in the same step. Plan accordingly.

Then, the very next day, DeepSeek dropped V4. Two preview models: V4-Pro at 1.6 trillion total parameters, 49 billion active, and V4-Flash at 284 billion total, 13 billion active. Both support a one-million-token context window. Both are open weights on Hugging Face, and both run on Huawei Ascend chips 678.

V4-Pro is now the largest open-weights model ever released, surpassing Kimi K2.6’s 1.1 trillion and GLM-5.1’s 754 billion, and more than doubling DeepSeek V3.2’s 685 billion from December. The pricing is the part that lands hardest — V4-Flash undercuts GPT-5.4 Nano at fourteen cents per million input and twenty-eight cents output. V4-Pro, the frontier-class model, is one dollar seventy-four in and three forty-eight out — about a third of GPT-5.4’s pricing and a sixth of GPT-5.5’s 6. The DeepSeek paper also reports a real efficiency win: at 1M context, V4-Pro uses 27% of the per-token FLOPs and 10% of the KV-cache footprint of V3.2.

DeepSeek’s own assessment, in their paper: V4-Pro “falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately three to six months.”

Cybin take — that is the most honest self-assessment from a frontier lab in recent memory. Three to six months trailing, at one-sixth the price, with open weights. The economics of the open-closed gap just changed.

Last week’s brief flagged DeepSeek V4 as an open thread and called the gap-to-frontier the line to watch. The thread is resolved: open weights are now within striking distance of GPT-5.4-class capability, at consumer-affordable inference cost, with a million-token context. Whether that translates to GPT-5.5-class is the next question, and DeepSeek’s own number — three to six months — is the time horizon to assume.

Now the third thing, which is interesting in a different way. Anthropic published a public engineering postmortem for Claude Code 910. They acknowledged three real regressions between March 4th and April 20th that explain the months of complaints about Claude Code feeling worse. First, they changed the default reasoning effort from high to medium to reduce latency, and reverted to xhigh on Opus 4.7 once the user reports stacked up. Second, a thinking-cache bug shipped on March 26th caused the system to clear stored reasoning on every turn instead of once per hour, which made the model appear forgetful, repetitive, and which also drained user budgets faster than expected. Fixed in version 2.1.101. Third, between April 16th and April 20th, a system-prompt instruction added “keep text between tool calls to twenty-five words or fewer,” and broader evaluation testing later revealed it dropped intelligence by three percent. Reverted on April 20th.

The going-forward commitments matter: broader internal testing using the exact public builds, multi-repository code-review tooling, stricter controls on system-prompt changes with per-model evaluations, and soak periods with gradual rollouts for any change that touches intelligence.

Cybin take — this is the engineering-culture story of the week, and it is not flashy. A frontier lab publishes a detailed regression timeline and ships process changes. That is the marker of a maturing operational discipline. Watch whether OpenAI and Google match the disclosure cadence.

A few items in the read tier without their own segments. OpenAI also shipped a Codex Superapp this week — Codex Automations, a plugins-and-skills system, and a deeper Workspace Agents integration 11. This is productization, not a frontier capability shift, but the Codex weekly active user count has reportedly hit four million, and the Accenture-PwC-Infosys partnership announcements suggest the enterprise distribution race is in active phase 12. Notably, OpenAI’s framing now treats Codex as the primary developer surface rather than as a feature.

From DeepMind, two items hold open from last week without major new development. Decoupled DiLoCo, the resilient distributed-training paper, still awaits independent reproduction 13. And Berkeley’s BAIR Gradient-based Planning for World Models at Longer Horizons paper — using gradient-based methods to plan over longer-horizon world-model rollouts — is the kind of paper that does not move benchmarks today but seeds capability shifts in eighteen months 14.

Jack Clark’s Import AI 454 has a useful framing on automated alignment research and a safety study of a Chinese model — worth reading once if you track alignment-research pipelines 15.

Cybin take — the meta-pattern this week is convergence at two altitudes. At the model layer, GPT-5.5 and DeepSeek V4 hit the same week, which means closed and open are now competing on the same release calendar. At the operations layer, Anthropic's postmortem and OpenAI's Codex superapp both signal that the labs are now competing on engineering polish, not just on capability. The frontier is bifurcating into capability races and reliability races in parallel.

Threads to watch into next week. The GPT-5.5 API rollout — once it opens generally, expect a wave of comparative benchmark posts and the first credible head-to-head against DeepSeek V4-Pro. Independent reproductions of DeepSeek V4-Pro’s frontier benchmarks, especially on coding and long-context tasks where the one-million context window is supposed to be agent-usable. Whether Kimi K2.6, the previous open-weights leader, gets a quick refresh response to V4. And the Anthropic postmortem’s promised process changes — soak periods and per-model evaluations — should show up as visibly slower release cadence on Claude Code over the next four-to-six weeks if they hold the line.

That’s the week. Stay sharp.