They Stopped Undercutting Claude. Now They Are Trying To Beat It.

CLAUDE STILL LEADS WHERE WE ACTUALLY WORK

Opus 5.5, released on September 22, 2026, leads on most coding benchmarks. It does so in Anthropic’s own table, in Google’s own run of the comparison, and on the independent board at Vals AI, which runs every model through a single harness. The ordering has held across all three sources.

The one place Argon looks ahead is human preference. It topped a preliminary Text Arena ranking at launch, though we could not check the live board.

Google announced Argon on September 30, first only to vetted cyber defenders through its Fairwind program. On October 13 Google confirmed that Argon is rolling out to AI Pro and AI Ultra subscribers in the Gemini app. According to press reports, an introductory API price rises on November 30, and we could not confirm Gemini CLI support. Argon is the challenger I would test first.

Meta released Muse Wonder and Muse Wonder Full on October 12, calling them its first frontier models. Meta built its open-model reputation on Llama, and both of these models are closed. Wonder costs $4 and $20 per million tokens, the same as Opus 5.5, and Meta says it is also on OpenRouter. Full is on the $50 and $200 Muse Code tiers, per Meta. In use it sounds rougher, per community reports of looping review-fix cycles, harness problems and 429 quota errors within minutes on the $200 plan.

ANTHROPIC ASKED EVERYONE TO SLOW DOWN, THEN SHIPPED THREE MODELS

Dario Amodei published “We Must Pace the Frontier” in September 2026. It proposes pacing the frontier, with a full pause only as the outermost option, which he expects to stay unlikely. One explainer summarizes it as “Training continues. Releases continue. What changes is the slope.” Press reported that Sam Altman said OpenAI would match the first commitment, embedded third-party evaluators, and that Elon Musk posted “Dario is right.”

Opus 5.5 arrived on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7, while OpenAI shipped GPT-6.1 Sol and Google shipped Argon on September 30.

The straightforward read is that progress has become hard to predict and the call is sincere. The cynical read is that the leader asks the field to slow down while it holds the lead. I do not know which read is right.

THE NEW RELEASES ARE SMALL, CHEAP AND SPECIALISED

At the frontier the open labs mostly promised. DeepSeek V4.1 Flash, released under the MIT licence in September 2026, is the main one that shipped weights, Mistral and Reflection only promised them, and the best open models sit well behind on the independent agentic-coding board. Below the frontier, open bases underpin most recent small releases.

Jev, from TypeSafe AI (September 15), returns typed decisions instead of prose. An aggregator counts at least six vendors shipping Jev-style models in about four weeks, which I hold loosely. Cloudflare’s Clef (October 1) is Apache 2.0, Jev-API compatible and runs on Workers AI, and Clef-omni (October 9) adds multimodal input. Microsoft-Decision-1 (October 9) matches Jev’s price on Foundry and OpenRouter. Clef-flash lists at $0.038 per million input tokens against Jev’s $0.042, output free. Cloudflare says Clef is faster, TypeSafe’s own tests still favour Jev, and both are vendor-run.

JetBrains released Mellum2.1 on October 8, a small coding-agent model open under Apache 2.0. It is a candidate local subagent for client code that cannot leave a client’s servers, though in JetBrains’ own tests an existing Qwen model still fixes more bugs.

Paper Instruments released Ultramarine 1 Flash on October 14, 2026, an RL post-train of Z.ai’s open GLM-5.3 Flash for finance, law and consulting work. The vendor claims near-Fable 5 results at about one-sixtieth the cost per task, on vendor-built tests that put it last on its own coding rows and compare it with older rivals. No independent evaluation exists yet, it is available only through its own API, and its weights are promised with no licence stated. It is not a tool for our coding work.

These releases compete on price, speed and narrow fit, which is what I would expect once the general-purpose frontier stops moving much.

THE INTELLIGENCE RACE IS RUNNING OUT OF YARDSTICKS

I expect model capability to level off, and I expect the other labs to reach Fable-class capability soon. The strongest case for this view is benchmark saturation. Frontier models already beat the average person on most tests we know how to write and, on a growing list of tests, arguably their makers. When the best models cluster at the top of a test, it stops separating them, and further gains are hard to show without a new yardstick.

Catch-up is already visible. Anthropic says Opus 5.5 matches Fable 5.1 “on most work” at about 40% less cost than Opus 5. Meta claims Muse Wonder is comparable to Fable 5.1, and Ultramarine is at best a signal, since its claims are the vendor’s own. If raw intelligence does level off, labs will compete on cost, speed, usage limits and the harness, and this issue already shows it in Muse priced exactly at Opus, OpenAI reworking Pro, Argon on an introductory price and the small fast models above.

I hold this view loosely. The people closest to these models are asking the industry to pace the frontier, which they would hardly do if they saw the slope flattening. A saturated test can also mean the test ran out before the capability did, and past plateau calls have been wrong. If I am right, value moves toward the harness and routing and away from any one model.

WHAT THIS MEANS FOR US

If models converge, a replaceable model matters more, which is still no reason to replace ours. An earlier issue said it this way: “The model should be swappable. That is not the same as worth swapping.”

Limits are already part of the race described above. Anthropic confirmed it raised five-hour limits on paid plans alongside Opus 5.5, with a one-time reset subscribers can bank. According to press reports, OpenAI cut its $200 Pro tier to 10 times Plus usage, with the old allowance kept until October 29. A small slice of our usage runs token-billed through OpenRouter, which is where challenger tests belong.

• Keep Claude for daily development. Nothing here changes the bar to switch.

• Test challengers through OpenRouter on a small spend, against internal repos only, or ask the AI Excellence team. Do not use company Gemini or ChatGPT plans, and check the listings first.

• If you test Argon, consider doing it before November 30, when the reported price increase lands.

• Judge every test on cost per completed task, in a real repo on a real ticket.

• Consider spending the banked Opus 5.5 reset on a heavy migration day.

• Watch the Argon and Muse harnesses, Gemini CLI support, OpenAI’s Pro limits after October 29, whether other labs follow the Amodei pacing commitments, and whether other labs reach Fable-class capability. AI Excellence team: trial one decision model, such as Jev, Clef or Microsoft-Decision-1, against Haiku 5.5 (October 7, 2026) with structured output before trusting any of them, and never let one auto-approve permission prompts in client repos.

Sources → Anthropic Opus 5.5  ·  Anthropic Haiku 5.5  ·  Dario Amodei  ·  Channel 4 News  ·  Computing for Geeks  ·  Google blog  ·  Khaleej Times  ·  OfficeChai  ·  HackerNoon  ·  Manifold  ·  TechCrunch  ·  Meta developer docs  ·  Hacker News  ·  MindStudio  ·  Vals AI  ·  Hugging Face  ·  ProPakistani  ·  TypeSafe AI  ·  Paper Instruments  ·  OrcaRouter  ·  Let's Data Science  ·  Cloudflare docs  ·  Cloudflare blog  ·  Pasquale Pillitteri  ·  Microsoft Command Line  ·  DataNorth  ·  NoCode MBA