AI Briefing · Thursday, October 8, 2026

The day in brief. Three threads today. First, small-model prices keep falling. Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 below 100k tokens but costs more above that, so the shape of a customer's workload now decides gross margin. Second, two board-seat companies shipped: Spellbook made its autonomous contract management generally available, and Cohere released open weights for a 3.35B multilingual reasoner. Read both with that bias in mind. Third, AI-math claims are running ahead of the evidence: Marcus says OpenAI's proof report omits its method. Diligence ask this week: gross margin per task at current prices, and at half.

Top of mind

• Small-model prices converge. Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 under 100k tokens; above that, Haiku charges $0.50/$2.50 against Luna's $0.20/$0.75 at 272k. 3
• Spellbook goes autonomous. Its contract management tool is now generally available. No customer, pricing or accuracy data and no critical pass, so treat it as a launch, not traction. 2
• Cohere's small multilingual reasoner. Cohere reports reasoning in the prompt's language over 93% of the time across 60 languages at 3.35B, with open weights. Self-reported; the accuracy cost is not quantified. 4
• AI math claims lack method. OpenAI released hundreds of proofs, but Marcus and Tao say the report lacks method detail. Mollick's 'superhuman' and Navier-Stokes claims are unverified. 1

Must read

1. OpenAI releases hundreds of AI math proofs; Marcus calls report too vague

Gary Marcus · Ethan Mollick (One Useful Thing) · 3 sources · source

• What happened: Gary Marcus and Terence Tao critique OpenAI's math benchmark claims (Gary Marcus), noting missing method details in their report. Ethan Mollick cites Scott Aaronson's blog post describing AI-generated math proofs as 'superhuman' (Ethan Mollick, One Useful Thing) but provides no verification; he also claims AI produced novel Navier-Stokes proofs without evidence (Ethan Mollick).
• Evidence check: OpenAI's report lacks method details per Marcus. Mollick's claims of superhuman proofs and novel math results lack verification or named systems.
• Editor's view: Don't accept founders citing these proofs as evidence of reasoning gains until OpenAI discloses method, compute and human assistance. Ask an industry expert how DeepMind would audit such a claim.
• Sources: Gary Marcus: Gary Marcus and Terence Tao critique OpenAI's math… · Ethan Mollick (One Useful Thing): Ethan Mollick shares early accounts of narrow superhuman… · Ethan Mollick (One Useful Thing): Ethan Mollick claims AI can produce novel mathematical…

2. Spellbook launches autonomous contract management tool

Scott Stevenson · source · 2026-10-07

• What happened: Spellbook has launched its autonomous contract management tool, now generally available to customers.
• Editor's view: Ask Spellbook for usage since general availability, error rates on autonomous actions, and whether autonomy changes pricing from per-seat to per-contract.

3. Claude Haiku 5.5 matches GPT-6 Luna pricing

Simon Willison · source · 2026-10-07

• What happened: Anthropic launched Claude Haiku 5.5 with $0.10/$0.50 input/output pricing up to 100k tokens, matching OpenAI's GPT-6 Luna price; beyond 100k tokens, Haiku costs $0.50/$2.50 vs Luna's $0.20/$0.75 at 272k tokens.
• Agenda: Anthropic's product announcement - self-promotional with pricing details
• Editor's view: Price parity turns model choice into a routing decision. Long-context portfolio workloads should benchmark Luna, and a startup whose only edge is per-token price has no moat.

4. Cohere Releases Tiny Aya L2-Thinker Model for Multilingual Reasoning

Cohere / Cohere Labs · source · 2026-10-06

• What happened: Cohere says Tiny Aya L2-Thinker (3.35B) reasons in the prompt language over 93% of the time across 60 languages (45 seen in reasoning training), with minimal accuracy loss except on competition-level math; weights and data released.
• Evidence check: Overstated. 93% in-language rate is plausible but self-reported; 'minimal accuracy loss' unquantified here, competition math excepted.
• Editor's view: Open weights mean the 93% figure can be checked independently; get per-language accuracy deltas before treating this as a sovereign-AI proof point.

5. Ethan Mollick predicts two AI-driven scientific revolutions

Ethan Mollick (One Useful Thing) · source · 2026-10-07

• What happened: Ethan Mollick (One Useful Thing) states that AI will cause two scientific revolutions: 1) re-evaluation of all published research through unanticipated methods, and 2) faster discovery rates than human absorption can keep up with.
• Agenda: Ethan Mollick is a researcher and educator who writes about AI's impact on work, not a lab staff or company spokesperson.
• Editor's view: Opinion with no evidence or timeline. Use it as a thesis frame for research-tooling startups, not as a basis for underwriting.

6. Copier Repair Tech Ethnography Highlights AI Work Replacement Complexity

Ethan Mollick (One Useful Thing) · source · 2026-10-07

• What happened: Ethan Mollick references a 1990s ethnography of copier repair technicians showing work complexity and improvisation, noting eventual tech-driven elimination.
• Editor's view: A useful counterweight to replacement pitches: ask agent startups what undocumented improvisation their customers' workers do, and how the product handles it.

In brief

• Ben Affleck discusses ML in film workflows (Simon Willison)

Open questions

• Will OpenAI publish method, compute and human involvement for its released math proofs? Shows whether the proofs reflect model capability or curated, assisted output. Settled by: A methods appendix, or independent verification by Tao or other mathematicians. 1
• Will Haiku 5.5's higher price above 100k tokens push long-context volume to GPT-6 Luna, or force an Anthropic price cut? Sets inference margins for portfolio apps handling long documents, including legal tools. Settled by: An Anthropic price change, or model usage-share data from routing providers. 3
• Do Spellbook's autonomous contract features drive paid expansion, or stay a demo feature? Tests whether legal AI moats go beyond orchestration on frontier models. Settled by: Attach rates, net revenue retention and autonomous-action error rates reported by Spellbook. 2

9 items analyzed from 5 sources · 3 filtered as off-topic · 0 more stories in the dashboard. Analysis: claude_code:default, local:qwen3:30b. Editor's view and day-in-brief are machine-written. Generated 2026-10-09 09:25 EDT.