Week of 2026-10-11

Summary

This update covers October 5 through October 11, 2026, following the previous update on October 4.

A quiet week by release standards and an unusual one by volume. On October 6 OpenAI posted 722 mathematics manuscripts written by an unreleased model, including claimed proofs of famous conjectures. Roughly four in ten have a main result checked by the Lean proof assistant; the rest carry OpenAI’s own warning that they may contain errors. On October 7 Anthropic shipped Claude Haiku 5.5, carrying the falling-price pattern from the flagship down to the cheapest tier.

The baseline remains moderate acceleration. The math release is the more interesting item and the less settled one: it is a large claim whose verification is still in progress. Nothing this week bears on recursive self-improvement or self-directed agency.

Key Developments

722 manuscripts, one model, and a verification queue

OpenAI published 722 machine-written mathematics manuscripts on October 6, including claimed resolutions of major open problems. A secondary-reporting estimate as of October 7 put the Lean-checked share at roughly 42%. The distinction matters: a formalized proof can be checked mechanically, while an unformalized manuscript still needs its reasoning verified.

Prompts and model identity were withheld. Scott Aaronson reportedly wrote that no human yet appeared to understand most of the proofs. Formalization and expert understanding are different checks, and neither the output volume nor that dated percentage establishes the significance of the results.

Interpretation: a Lean check establishes that the logic is valid, not that the problem was stated correctly or that the result is significant, so a verified proof of a mis-stated problem would still pass. That makes the 42% figure good news about machine-checkability and neutral about importance. It also continues the pattern the baseline recorded in August, when Astra’s ten problems and Claude’s zeta-function bound arrived: the problems were chosen by humans, and the system did the work. The withheld prompts mean outsiders cannot tell how much human steering sat inside “a single prompt.”

Speculation: if even a modest fraction survives review, the registry’s 50-problems-by-2030 threshold would be reached years early. If most of the unformalized remainder fails, the episode will read as a reminder that generation is cheap and checking is not. The next few weeks of expert reading will settle which.

Sources: openai-math-manuscripts-2026

The falling price reaches the cheapest tier

On October 7 Anthropic released Claude Haiku 5.5 at an input-token price about a tenth of Haiku 4.5’s for prompts up to 100K tokens. Anthropic reports roughly 75% lower average running cost, a different measure that accounts for differences including tokenization. Independent scores put it level with rivals’ cheap tiers; the rest of the benchmark detail is vendor-reported.

The pattern is the one the baseline has followed all year: the Sonnet 5, Opus 5 and Opus 5.5 releases lowered the price of competent autonomous work at progressively higher tiers, and this one lowers it at the bottom. The practical consequence is for high-volume work such as classification and support, where price per task decides whether automation is worth doing at all. It is a price and cadence datapoint, not a capability ceiling move.

Sources: anthropic-haiku-55-2026

Baseline Impact

Updated:

• Section 2, release cadence and price. Extended the falling-price passage with Haiku 5.5, noting the tokenizer offset and the thin independent evaluation.
• Section 8, verifier/mathematics passage. Added the October 6 OpenAI manuscript release as a third result alongside Astra’s ten problems and Claude’s zeta-function bound, with the partial Lean-verification status and the point that Lean validates logic rather than significance.
• Section 2 header, Section 9 / Last updated. Moved the snapshot label and date stamps from “late September” to “early October” and bumped the date line to 2026-10-11.

No change:

• Moderate acceleration remains the central scenario.
• No evidence of recursive self-improvement or self-directed agents. The problems in the math release were human-posed.
• No status changes in the prediction registry (see below).

Scenario Impact

Moderate acceleration. Unchanged as the central case. Falling prices at every tier and a large but unverified burst of machine-generated mathematics both fit steady, uneven advance.

High acceleration. Mildly supportive, conditional. If a meaningful share of the 722 manuscripts holds up, AI-driven research output would be scaling faster than this path’s near-term milestones assumed. That is a conditional, not an observation, and the week’s evidence does not yet support it.

Low acceleration / regulated path. Roughly neutral. Nothing this week touched governance, and the math release cuts neither way for this path until its verification resolves.

Risks and Opportunities

Risks:

• Verification capacity is now the bottleneck in at least one field. Hundreds of unreviewed manuscripts compete for the attention of a small number of experts who can judge them, and an unreadable proof is hard to referee even when it is correct.
• Withheld prompts and model identity reduce reproducibility, which is the usual route by which an impressive claim becomes an accepted one.

Opportunities:

• Formal verification offers a cheap, objective filter. If the Lean-checked share rises, mathematics becomes the clearest field where machine output can be trusted without trusting the machine.
• Models priced at $0.10 per million input tokens for prompts up to 100K tokens widen the set of high-volume tasks worth automating.

Required Baseline Changes

Applied in this run: Section 2 (Haiku 5.5 added to the price passage; snapshot renamed “Early October 2026”), Section 8 (OpenAI manuscripts added to the mathematics passage), and Section 9 date stamps. Two sources were added; no new prediction or theory was needed.

Registry: no status changes. The manuscripts are not yet countable under jf-math-acceleration-16, which requires a long-open problem, a machine-verified proof and credible reporting; the human-posed problems leave jf-math-direction-17 unmet. Both received evidence notes.

Watch Next

• Independent expert assessment of the Unique Games Conjecture and Dirichlet L-function claims, and whether the Lean-checked share rises above roughly 42%.
• Whether OpenAI names the model, discloses prompts, or ships it — the move that would turn a claim into something others can reproduce.
• Whether individual results can be counted under jf-math-acceleration-16.