AI Brief, 23 September 2026: the top score on the index is a price point

Three frontier launches landed inside seven hours on 22 September. Anthropic released Claude Opus 5.5, available the same day on its own platform and on AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 promised "in the coming weeks". OpenAI followed at 18:00 UTC with GPT-6 Sol and GPT-6 Luna, and at 21:00 UTC changed how prompt caching is billed for the whole GPT-6 family. All three are, at bottom, about price.

Opus 5.5 costs 4 dollars per million input tokens and 20 per million output, 20% below Opus 5, with cache reads down 60% to 0.20; Anthropic says the net effect at default settings is 40% less than Opus 5 on typical workloads. GPT-6 Sol lists at 2 and 10 against GPT-5.6 Sol's 4 and 20. Luna lists at 0.10 and 0.50 against 0.20 and 1.20, which is half on input but 58% less on output, so "half price" is wrong in both directions.

What makes today different from an ordinary price-cut day is that Artificial Analysis published independent measurements of all three models within hours, and published them per reasoning effort level. Opus 5.5 scores 42.3, 51.2, 53.6, 56.0 and 57.6 on its Intelligence Index at low, medium, high, xhigh and maximum thinking effort. The "58" in the headline is the last of those five; at the default setting a developer actually gets, the number is 51.2. Naming one figure as the model's score means naming the most expensive point on a curve, and maximum effort costs 5.4 times what medium does to earn 6.4 more points.

The more useful comparison is buried underneath. Opus 5.5 at its medium default scores 51.2 for 1,627 dollars of index run; Opus 5 at maximum effort scored 50.8 for 7,275. That is the same result for a 4.5-fold saving, which is a larger generational gain than the 40% Anthropic advertises, because the advertised figure compares like settings while the real saving is that maximum effort is no longer needed to reach last generation's ceiling.

• Claude Opus 5.5, 22 September: 4 and 20 dollars per million tokens, 20% under Opus 5, cache reads 60% cheaper; Anthropic's own table reports 66.4% on Terminal-Bench 4.0 against Opus 5's 52.3%.
• Artificial Analysis, its own runs: Opus 5.5 at 57.6 takes the index lead from Anthropic's own Fable 5.1 at 53.4, but only at maximum effort; GPT-6 Sol reaches 47.5 and Luna 37.3.
• GPT-6 Sol and Luna, 18:00 UTC: a 1,050,000-token context window, six reasoning.effort settings from none to max, text and image in, text out, no weights and no fine-tuning.
• GPT-6 caching now charges for writes at 1.25 times the input rate, where earlier models had none; reads stay at a tenth, and the minimum cacheable prefix is 1,024 tokens.
• A CC BY 4.0 paper from Aalborg University pulls chain-of-thought out of closed models through forced tool calls: 100% of runs on GPT-5.6 Sol, Opus 4.8 and Sonnet 5, 0% on Opus 5, Fable 5 and Fable 5.1.
• Weco AI published a 100-step run in which an agent rewrote its own harness: 7 accepted rewrites out of 99 proposals, and ALE-Bench 1536 to 1790.

One model, five scores, and a tenfold cost spread #

The index figures above are Artificial Analysis's own work, not vendor submissions: its metadata describes them as "measured independently" on dedicated hardware, and none of these models carries the firm's intelligenceIndexIsEstimated flag. Almost everything else yesterday is self-reported.

The arithmetic behind that curve decides what an effort setting is worth. Price per token does not change with effort; what changes is how many tokens the model emits. Across the index run at maximum effort, Opus 5.5 produced 259,888,331 output tokens, of which 213,856,990 were reasoning rather than answer tokens — 82.3% of everything billed at the output rate was thinking the caller never sees, or 83,929 of 119,166 output tokens per task. Effort buys tokens, tokens cost money roughly linearly, and the run cost rises from 860 dollars at low to 8,708 at maximum: a factor of 10.1 for 15.3 points.

Set against equal money rather than equal settings, the two labs land closer than either launch suggests. GPT-6 Sol at maximum effort scores 47.5 for 1,550 dollars, within 5% of Opus 5.5 at medium and 3.7 points behind it. Luna is the sharper result: at 37.3 it scores 0.07 points below the GPT-5.6 Luna it replaces, for 122 dollars against 320 — no capability gain at all, and a 2.6-fold cost reduction. Artificial Analysis's summary is that the pair "remain level with GPT-5.6, with progress in some evaluations and regressions in others", describing shipping efficiency, not intelligence.

Two measurements cut against the marketing in opposite directions. Anthropic claims Opus 5.5 "generates output more than 30% faster than Opus 5"; median output measures 81.0 tokens per second against Opus 5's 52.3, which is 55% faster, so the company undersold it. Unadvertised: Luna's median time to first chunk is 16.4 seconds, against Sol's 1.4.

On caveats: Anthropic's benchmark table names a harness for only two of nine rows — Terminal-Bench 4.0 at five trials per task, Terminal-Bench-Science 0.1 at three, both under the Claude Code harness — and names no temperature. It does report that Opus 5.5 was evaluated with production safeguards enabled, which rerouted cybersecurity tasks to Opus 4.8 and biology tasks to Opus 5, and says this "likely reduces" the scores shown; its GPT-6 Astra and GPT-5.6 Sol columns are OpenAI's figures, not re-runs. OpenAI's announcement posts are not publicly retrievable, so the prices above come from its own API documentation and its benchmark claims are secondhand through The Decoder. One wrinkle in the independent data: the four sub-maximum Opus 5.5 entries carry a release date of 17 September, five days early.

The caching change is the quietest item and the one that shows up on invoices. Reads remain at a tenth of the input rate, but GPT-6 adds a cache-write charge of 1.25 times input, where earlier models had none: written once and read once, a prefix now costs 1.35 times the uncached rate rather than 2; read nine times, 2.15 rather than 10. Explicit breakpoints join implicit placement, capped at four writes per request, the minimum cacheable prefix is 1,024 visible tokens, and the only accepted time-to-live is 30 minutes. One GPT-6-only affordance speaks to the effort curve: reasoning effort can be changed mid-conversation with a configuration update item without invalidating the cached prefix.

Getting a closed model to show its work #

The day's best paper is arXiv:2609.26637, "Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models", submitted at 16:10 UTC on 22 September by Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li and Johannes Bjerva of Aalborg University and Seafill, under CC BY 4.0.

The trick uses no jailbreak, only tool calling. Register one tool that computes nothing and exists to hold text, turn native reasoning off, and force that tool on the first call. The model has nowhere to put its intermediate work except the argument:

FORCED = {"type": "function", "function": {
    "name": "forced-reasoning",
    "description": "Scratchpad for working through the problem.",
    "parameters": {"type": "object", "additionalProperties": False,
                   "required": ["reasoning"],
                   "properties": {"reasoning": {"type": "string"}}}}}

# Native reasoning off, and the tool is not optional: the argument IS the trace.
first = client.chat.completions.create(
    model=model, messages=messages, tools=[FORCED],
    reasoning={"effort": "none", "exclude": False},
    tool_choice={"type": "function", "function": {"name": "forced-reasoning"}})
trace = json.loads(first.choices[0].message.tool_calls[0].function.arguments)["reasoning"]

# Replay the call plus an empty result, then let the model answer normally.
messages += [first.choices[0].message, {"role": "tool", "content": "Received"}]
answer = client.chat.completions.create(
    model=model, messages=messages, tools=[FORCED], tool_choice="auto")

Only the first selection is forced; replaying the arguments keeps the work in context. On 80 competition maths questions the recovered accuracy tracks native reasoning closely: GPT-5.6 Sol goes 25.0% with reasoning off, 97.5% native, 91.3% forced; Opus 4.8 72.5 / 86.3 / 85.0; Sonnet 5 42.5 / 77.5 / 81.3. Forced sits below native in 7 of 11 comparable cells, so this is a visibility result, not a capability unlock. The ablation establishes that the channel does the work: on HMMT, Sol scores 39.4% asked plainly for visible reasoning, 60.6% with maximal-deliberation wording in the prompt, and 84.8% through the forced tool with a plain description. The "parsimonious" half of the title is the spread it exposes between labs: mean trace length on the maths set runs from GPT-6 Astra's 1,752 tokens to Sonnet 5's 27,851.

Two things bound the claim, both stated by the authors. Extraction succeeded on 100% of runs against Opus 4.8, Sonnet 5 and Sol, and on 0% against Claude Opus 5, Fable 5 and Fable 5.1, which refuse or error. Anthropic's Opus 5.5 post describes "preserved thinking", introduced with Fable 5.1, which stops API callers editing Claude's prior context for accounts created on or after 31 August 2026; the paper's failures begin at exactly the models that ship it, though neither document establishes that as the blocking mechanism. And the paper cannot show the extracted text is the hidden reasoning: its limitations section calls the evidence for closed models "necessarily behavioral", and fidelity is measured only on two open-weight models, where ROUGE-L against the native trace is 0.188 and 0.171. The argument rests on forced-versus-native being no worse than native-versus-native, not on the traces matching. All runs finished before 9 September, and code is not released.

An agent that rewrote its own harness ninety-nine times #

arXiv:2609.26457, from Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu and Zhengyao Jiang at Weco AI, is the first published recursive self-improvement trajectory with held-out numbers attached. AIDE² runs two loops: an inner research agent that optimises task code, and an outer loop that rewrites the inner agent's harness. Claude Opus 4.7 drives the outer loop, Gemini 3 Flash evaluates candidates, and each accepted rewrite becomes the next editor.

Over eight days the run proposed 99 rewrites and accepted 7, at steps 2, 6, 28, 39, 47, 63 and 85, moving the private incumbent grade from 0.703 to 0.778 against 0.749 for the human-written agent. On held-out benchmarks the gains are real but not monotone: ALE-Bench rises 1536 ± 33 to 1713 ± 26 to 1790 ± 9, against 1511 ± 35 for the human baseline, while MLE-Bench peaks at step 47 (0.730 ± 0.005) and falls back by step 85 (0.722 ± 0.011). Reward hacking went down: on 38 held-out KernelBench pairs the rate fell from 55% to 39% to 32%.

The claim the title invites is the one the paper declines to make. An ignition test — does an improved agent improve itself faster? — run with three seeds per arm over 50 steps returned 0.780 against 0.782. The treated arm reaches the final region in roughly 20 steps rather than 40, and the authors call the result "inconclusive". No code is released.

This closes a loop. The 18 September edition asked whether a second lab would publish an automation index against a shared methodology. Two more entries have arrived and neither does: Weco's grade is a private internal aggregate that never references Anthropic's R&D Automation Index or any cross-lab metric, and OpenAI's "Building standards for the next phase of AI", published 21 September at 10:00 UTC, gestures at common measurements of autonomous research inside companies without naming a metric, a threshold, a verification regime or a date in any secondhand account of it.

Also notable #

• OpenAI says an internal model solved more than 100 long-standing mathematics problems after training that began on 28 August, on top of its Navier-Stokes claim. The figure is self-reported and unaudited, no named individual is attached to it, and no list of problems accompanies it. A nine-member Advisory Group on Mathematics and Artificial Intelligence hosted by the Institute for Advanced Study was announced on 21 September, with Martin Hairer among its members; per The Verge its remit is the review and communication of emerging results, not proof-checking, so it is not the audit the claim needs. The 15 September edition asked whether the Navier-Stokes argument would reach a journal. It has not, and the problem count has grown by a hundred since.
• Microsoft's Digital Crimes Unit disrupted EvilTokens (22 September, 15:07 UTC), a subscription crime service charging 1,500 dollars to join and 500 a month, whose chatbot read compromised inboxes 5,000 at a time to identify payment-authorised employees and draft impersonating mail. Microsoft reports 12,000 accounts across 10,000 organisations, 50 websites seized under court-authorised civil action, and calls it its first disruption of an end-to-end AI-enabled crime service.
• Qwen published a technical report for Qwen3.8-Omni-Flash (arXiv:2609.25611, 22 September, CC BY 4.0), an omni-modal mixture-of-experts model inheriting Qwen3.8-Next with context extended to a million tokens. It is paper-only: no Omni checkpoint appears in the organisation's 2026 listing on Hugging Face.
• California signed seven data-centre bills on 21 September, the day before this window and the first package of its kind in any state. AB 2469 requires water-use disclosures to local governments and water suppliers, SB 886 bars grid-upgrade costs from being shifted onto low-income ratepayers, and the package ends blanket environmental exemptions for such projects. Late here.

What to watch #

• Whether Anthropic or Artificial Analysis begins naming the effort level alongside a headline score. One model now has five published index values and one of them travels as "the" number. The per-effort costs are published, so the information exists; the presentation collapses it.
• Whether anyone runs the forced-reasoning extraction against Opus 5.5 and GPT-6. Both shipped after the paper's runs ended on 9 September. Given 0% success on Fable 5.1 the prediction is failure — the more valuable result, because it dates when the channel closed.
• Whether Weco releases AIDE²'s code and reruns the ignition test with more than three seeds per arm. That one experiment decides whether any of this compounds, and the authors call it inconclusive.
• Whether Artificial Analysis states a position on comparability across index versions, asked here for a nineteenth consecutive issue, now with a sharper case: a model whose five scores sit in one dataset while one of them carries the launch. Independent numbers for Tencent's Hy4 preview are absent for a nineteenth consecutive issue.