Claude Opus 5.5 Found AttnRes-Canon: What Four Days of Architecture Search Actually Proved

On 6 November 2026, a research group called IAAI published an Arena v1 run record naming Claude Opus 5.5 as the agent that designed a Transformer architecture called AttnRes-Canon, plus a checkpoint, a replayable trace and a ten-minute write-up by Yi Ge explaining how it got there. Claude Opus 5.5 itself is not new — A​nthropic shipped it on 22 September 2026 at $4 per million input and $20 per million output tokens, and nothing about that model's price, context window or published benchmarks changed this month. What is new is dated within the last week: a public, checkable record of a frontier model running an entire architecture search on its own, including the branches that failed, and a downloadable model that came out of the other end. That is a narrower claim than the timeline suggests, and it is more interesting than a benchmark row.

The record is also the reason this is a "here is what is actually established" piece rather than a launch post. The arena labels its own entry self-reported. There is one seed. Nobody outside the operator has reproduced the number, and at the time of writing the Hugging Face checkpoint had been downloaded zero times. Read the sections below with that in mind — not as reasons to dismiss the result, but as the exact shape of the evidence.

What AttnRes-Canon actually is

Two existing ideas, combined, plus one thing neither of them does alone.

The first idea is Attention Residuals, from a K​imi Team paper posted to arXiv in March 2026 (2603.15031). Standard PreNorm residual connections add every layer's output with a fixed weight of one, so the hidden state grows with depth and each layer's contribution gets diluted. Attention Residuals replaces that fixed sum with softmax attention over earlier layers, so the mixing weights depend on the input. The cheap version, Block AttnRes, groups layers into blocks and attends over block summaries; the expensive version lets every layer read every earlier layer directly. In the arena's baseline, 23 layers sit in 8 groups, so a layer can read back the sum of an earlier group's outputs — chapter summaries, in the author's image, not the chapters.

The second is Canon layers, from Zeyuan Allen-Zhu's "Physics of Language Models, Part 4.1" (arXiv:2512.17351, December 2025 with a July 2026 revision). A Canon layer is a small width-4 convolution: while processing the current token it also mixes in the three tokens immediately before it.

AttnRes-Canon is Full Attention Residuals — 23 groups instead of 8, every layer reading any earlier layer on its own — with Canon layers at four positions: A at the input to attention and C at the input to the MLP, B on the query, key and value projections, and D on the MLP hidden layer. One component carries information along depth, the other along the sequence. The combined gain came in slightly above the sum of the two gains measured separately, which is the only genuinely new architectural claim in the report.

• Depth component — Full AttnRes: 23 groups, each layer reads earlier layers individually rather than through block summaries

• Sequence component — Canon layers at A, B, C and D: a width-4 causal convolution over each token and its three predecessors

• Cost — 1.66× the time per training update of Block AttnRes, the parent it beat

If that last line sounds like an afterthought, it is the one the search never priced in, and we will come back to it.

The ruler came before the search

The arena's rules are deliberately boring. Training data, 46,875 updates and FLOPs per token (within ±1%) are fixed; only the architecture may change; scoring is CORE-22 — the mean over 22 tasks, higher better — plus validation bits-per-byte, lower better; and during the search the agent may look only at training loss and dev-split BPB. A single full run costs more than 1,400 GPU-hours, so no search can afford to train a dozen ideas to completion. Everything depends on whether a short run ranks ideas the way a long one would.

So the first ~152 GPU-hours went on calibrating that ruler rather than on the problem. The obvious shortcut — squeeze the whole learning-rate schedule into 512 updates — turned out to misrank systematically: trained on 11 methods whose full results were already known, it produced a Spearman correlation of 0.40 against the final ordering, against a bar of 0.85 set in advance. The errors were not noise. All four looped architectures fell in the ranking and all four linear-attention architectures rose, because cutting the learning rate early flatters fast-converging models and amputates slow starters.

Taking a prefix of the real schedule instead — same schedule, stopped early, then carrying on with a real full run — reached 0.72 at 500 updates and 0.92 at 2,000. That is what fixed the protocol: 512 updates is only a floor for killing candidates that clearly failed to train, and every ranking decision is made at update 2,000, about 4% of a full run. The same calibration produced the noise floor: two runs of the same seed differed by about 0.0004 in dev BPB, so a candidate had to beat its parent by more than 0.001 to count as a win.

That is a better experimental design than most human architecture papers bother with, and it is worth saying plainly that the design, not the model, is what makes the result legible.

Three rounds, twelve candidates, one mirage

Round one tried five single changes on top of Block AttnRes, each carrying a written, refutable hypothesis. None cleared the −0.0010 line at update 2,000. Full AttnRes came closest at −0.0008; Canon at positions A and C reached −0.0006. A gated-attention variant led the whole field at update 512 by −0.0027 and finished behind at +0.0006 — the report calls it what it was, a mirage, and the same pattern recurs.

Round two combined the near-misses and added two mixture-of-experts variants at matched compute. Three candidates cleared the line: Canon at all four positions (−0.0021, 1.47× step time), Full AttnRes plus Canon-A/C (−0.0018, 1.30×), and Full AttnRes plus the gate (−0.0010, just over). The gate contributed almost nothing wherever it was added, while both other changes lowered dev BPB on every parent they were tried on. The two MoE variants — including the equal-compute configuration the agent had rated highest — finished last and second-last.

And the ranking flipped hard: the top four candidates at update 512 finished in exactly the reverse order at update 2,000.

Round three is where the search earns its keep. Rather than trust the update-512 ordering a second time, it waited for the round-two runs to reach update 1,536. Two candidates then shared Canon-A/C as their parent, one taking a step toward Full AttnRes and one adding Canon positions B and D; both sat about 0.002 below that shared parent. With budget for one more candidate, the agent merged the two instead of waiting out the last few hundred updates. The merged model reached a dev BPB of 0.7965 — 0.0038 below Block AttnRes and first in the whole search — at 1.66× the step time.

Total: 12 candidates, 3 rounds, 550 GPU-hours of search.

The final run, and how much of the lead survived

Short runs see only the flat part of the learning-rate schedule. Whether an early lead survives the decay has to be measured, and this is the part of the report that deserves the most attention.

AttnRes-Canon was ported into a formal method plugin and trained from scratch for 49.152B tokens on 32 H20 GPUs, single seed.

• CORE-22 — AttnRes-Canon 35.99 vs mHC 35.33 vs Block AttnRes 34.00 vs Dense Baseline 33.10

• Validation BPB — AttnRes-Canon 0.6605 vs mHC 0.6617 vs Block AttnRes 0.6630 vs Dense Baseline 0.6672

• Leaderboard position — first of 38 published Arena v1 results on both metrics, per the arena's own board

• Compute — 152 GPU-hours exploration, 550 search, 1,433 for the final run, 2,135 in total

First on both metrics, and the margin is not enormous: 2.0 CORE points over the method it was built on, 0.0025 BPB. More useful than the endpoint is the trajectory. The gap to Block AttnRes was 0.0037 at update 2,000, widened to 0.0058 while the learning rate held constant, and then decayed to 0.0024 by the end — more than half the advantage lost in the tail. The report pairs it with an outside case: a Dense Baseline with the gate added led by 0.0012 at update 2,000, comfortably past this search's own winning line, and finished 0.0002 behind. That is the honest limit of the whole method, stated by the people who ran it: an early lead buys you a hypothesis, not a result.

What is verified, and what you are taking on trust

The run record is unusually candid about its own provenance, and the candour is the reason to trust the process description more than the headline number.

• Independent reproduction — none. The arena labels the entry "self-reported" and the footer of the write-up says "Arena v1 · single-seed results"

• Seeds — one. The report lists "never ran a second seed" among the agent's four failures

• Human involvement — every stage's plan, rules and GPU budget was approved by the operator, whose recorded role states they chose no candidate and wrote no method code

• Review — the record lists a second model, GPT-6 Astra at high reasoning effort, reviewing the stage-three code and launch scripts before any GPU job, not choosing architectures

• Reproducibility material — a public MIT-licensed code path for the method, a checkpoint and logs on Hugging Face, and a trace intended to replay every instruction and result; the checkpoint had zero downloads when this was written

• Selection signal — dev-split BPB only; the agent did not see CORE-22 while choosing

That last item matters more than it looks. The agent could read the published results of other methods but picked candidates using only the dev split, which is the difference between searching and peeking.

The failure list is the other half of the value. Four days of context, compacted seven times, with candidate notes written to files so the thread could be resumed. The agent planned to make its round-three decision at update 1,536 but set reminders only at 512 and 2,000, so it stopped in the gap and resumed only when a human nudged it. It rated equal-compute mixture-of-experts highest and both variants finished last. It confused run-to-run variation with seed-to-seed variation. And it optimised quality with no constraint on speed, which is how a 1.66× slower model won a compute-matched contest on quality alone.

None of those are exotic. They are the failure modes anyone running a long agent loop already knows: a plan without a timer, a prior that does not survive contact, a variance model that is not the variance you have, and an objective function with a hole in it. What is new is seeing them itemised by the people who watched them happen, with the trace on GitHub.

What this changes for someone choosing a model

Nothing about Claude Opus 5.5's rate card, context or benchmarks. The useful shift is narrower and more practical: this is one of the first public, replayable demonstrations that a general-purpose frontier model, given a fixed evaluation harness and human budget sign-off, can run a multi-round experimental search and produce something a team would actually train.

That has a specific consequence for how you would use it. If you are considering pointing an agent at a search problem of your own, the transferable part of the report is the protocol, not the architecture: calibrate your cheap proxy against known full results before spending a single GPU-hour, set a noise floor and a winning threshold in advance from repeated runs, change one thing per candidate against a stated parent, and decide wins at a horizon you have shown to be predictive. Those four rules cost about 7% of the total compute here and are the reason the other 93% was not wasted.

The architectural result is harder to act on. AttnRes-Canon is a research method, not a served model. It exists as a method plugin and a 49.152B-token checkpoint, single seed, and anyone wanting to build on it is starting from a reproduction attempt. Treat the CORE-22 figure as a lead worth a second seed, not as a settled property of attention residuals.

Where OrcaRouter fits is the part of the workflow that surrounds the search rather than the search itself. Claude Opus 5.5 is on our catalogue at A​nthropic's published list price passed through with 0% markup — $4.00 per million input, $20.00 per million output, $0.20 cache reads, 1M-token context — and the models a search loop typically gets split across sit behind the same key: Claude Fable 5.2 at $8.00/$40.00 since 9 November, GPT-6 Astra at $10.00/$50.00 up to 272K tokens and $20.00/$75.00 above it, Claude Opus 5 at $5.00/$25.00. A four-day research session that compacts its context seven times is a long, expensive, failure-prone conversation; running it against one endpoint with automatic failover means a provider hiccup costs a retry rather than a night, and the routing DSL lets you put a cheaper model on the routine analysis and keep Claude Opus 5.5 for the decisions that set the next round.

What to watch next

Three things would move this from an interesting self-report to a result you can build on. A second seed, which the report says did not happen and which would tell you whether 0.0025 BPB is real. A speed-constrained rerun, since a 1.66× step time is a third of the value if you care about throughput per GPU-hour rather than per token. And Arena v2, which the authors say will run inside a sealed container with a fixed compute allowance so the agent cannot see the formal validation set or CORE-22 at all — that is the version where "an agent did the research" stops needing a human in the loop to be credible.

Until then, the fair summary is this: an agent given a well-built ruler found a real, modest, reproducible-in-principle architecture improvement in four days for about 2,100 GPU-hours, and wrote down everything it got wrong along the way. The last part is the part the field usually leaves out.