AI Briefing · Tuesday, October 13, 2026

The day in brief. Price, not top-end capability, is today's story. GPT-6.1 Sol scores one point below Astra at 22% of the per-task cost. Mistral Large 4 matches peers' scores at over 4x their per-task cost. Labs are also protecting their outputs: Anthropic says distillation harvesting is down 90–99%, but nobody outside has checked. Treat the claimed $20B cut to OpenAI's revenue forecast as rumour. Ask in diligence this week: if your model bill fell 75%, what happens to gross margin and pricing, and does the product depend on distilling a closed model?

Top of mind

• Near-frontier at a fifth of the price. GPT-6.1 Sol scores 52 to Astra's 53 on Artificial Analysis, at $0.72 vs $3.26 per task. That is benchmarks only, but it reprices every margin model built on Astra. 3
• Mistral prices above its tier. Mistral Large 4 scores 38 on the AA Index at $1.13 per task, over 4x GLM and DeepSeek. Open weights are planned for late October. 4
• Distillation may get harder. Anthropic says new defenses cut distillation harvesting 90–99%. The evidence is internal tests with no published method. Startups training on Claude outputs should check their exposure. 5
• OpenAI turmoil, both stories contested. Three safety researchers were fired, and both sides' accounts are on record. Gary Marcus's claimed $20B forecast cut and 2027 IPO have no sources. 2 1

Must read

1. Marcus says OpenAI cut revenue forecast $20B, pushed IPO to 2027

Marcus on AI (Gary Marcus) · source

• What happened: Gary Marcus claims OpenAI cut revenue forecast by $20B (Marcus on AI), pushed IPO to 2027, and Anthropic is 'supposedly' IPOing next month without S-1 (Gary Marcus). An unnamed U.S. appeals court ruling 'undermined' generative AI companies (Marcus on AI).
• Critical pass: Overstated. Unverified claims: $20B forecast cut, unnamed court ruling, and Anthropic's unconfirmed IPO timeline lack sources or evidence.
• Our read: We treat the $20B cut as unconfirmed until a second outlet reports it with documents. For companies counting on OpenAI price cuts or an OpenAI-led exit market, nothing changes this week; a confirmed S-1 from either lab would.
• Sources: Marcus on AI (Gary Marcus): Gary Marcus warns of frontier labs' WeWork moment · Gary Marcus: Gary Marcus Questions AI Valuations Amid Market Shifts

2. OpenAI Fires Three Safety Researchers

Don't Worry About the Vase (Zvi Mowshowitz) · source · 2026-10-12

• What happened: Three OpenAI safety researchers (Tomek Korbak, Mikita Balesni, Jasmine Wang) were fired for 'mishandling sensitive company information,' though they claim it was due to safety advocacy. They published an open letter stating the firing was pretextual.
• Agenda: Rationalist/safety-concerned source skeptical of lab PR; strong priors against OpenAI
• Critical pass: Holds. The firings and both sides' accounts are well sourced. Why they were fired is still unknown, and Zvi admits it.
• Our read: We read this less as a governance signal than as a talent signal: safety-minded researchers now have a reason to look elsewhere. Seed companies hiring for alignment or evals have an unusual recruiting window.

3. OpenAI's GPT-6.1 Sol Nearly Matches Astra at 20% Cost

The Batch (DeepLearning.AI) · source · from Fri Oct 9

• What happened: GPT-6.1 Sol scores 52 on Artificial Analysis Intelligence Index vs GPT-6 Astra's 53, with $0.72/task cost vs Astra's $3.26; priced at $2/$10 per million tokens (vs Astra's $10/$50).
• Agenda: Andrew Ng's The Batch (DeepLearning.AI) reporting on OpenAI product announcement; no conflict.
• Critical pass: Holds. Independent Artificial Analysis data backs it: Sol is 1 point below Astra at 22% of per-task cost, on benchmarks only.
• Our read: This helps application-layer companies most: a model one point off the top at 22% of the task cost turns many marginal unit economics positive. Ask every API-heavy portfolio company to rerun gross margin on Sol pricing this week.
• The market view: OpenAI is undercutting its own flagship: $2/$10 per million tokens against $10/$50. A lab competing on price with itself squeezes lab margins, not only app-layer pricing.

4. Mistral Large 4 scores competitively with US/China models at premium pricing

The Batch (DeepLearning.AI) · source · from Wed Oct 7

• What happened: Mistral Large 4 (1T total, 49B active) is in Research Public Preview. Open weights are planned for late October. It scores 38 on the AA Index and costs $1.36/$4.18 per M tokens. Cost per task is $1.13, or $0.57 discounted, vs $0.25 (GLM) and $0.27 (DeepSeek).
• Critical pass: Holds. Third-party scores back it: about on par with GPT-6 Luna and DeepSeek Flash, but over 4x their cost per task.
• Our read: Mistral is betting buyers will pay 4x GLM's and DeepSeek's per-task cost for sovereignty and open weights; if that premium doesn't hold, Cohere faces the same test.
• The policy view: Open weights from a European lab in late October give non-US buyers a sovereign option. For them, per-task cost against GLM or DeepSeek is not the deciding comparison.

5. Anthropic claims new defenses cut distillation data harvesting 90–99%

Anthropic · Techmeme · SCRUTINY · 2 sources · source

• What happened: Anthropic claims new defenses reduce distillation data harvesting by 90-99%, citing internal tests (The Batch). Fable 5.1 and Opus 5.5 include preserved thinking; an older defense scored 71.8% vs 80.4% on SWE-bench Verified (Techmeme).
• Critical pass: Unsupported. Anthropic's self-reported 90-99% reduction lacks method or external verification, with the SWE-bench test using outdated defenses.
• Our read: We take this as a warning to any startup whose training leans on Claude outputs, whatever the real reduction rate. Ask founders where their fine-tuning data came from and whether the plan survives losing frontier-model distillation.
• The market view: Distillation defenses protect lab margins against cheap clones. If they work, closed models gain pricing power, and the gap with low-cost open models may persist.
• Sources: Anthropic: Anthropic claims 90-99% reduction in distillation attacks · anthropic.com via Techmeme: Anthropic reports 90-99% reduction in data harvesting via…

Worth knowing

• 6. JPMorgan and Goldman lead $5B loan sale for AI startup Volta at 11% yield. JPMorgan and Goldman Sachs led a $5 billion loan sale for AI startup Volta with an ~11% yield, underwriting only $1.3 billion to limit exposure. (bloomberg.com via Techmeme)
• 7. Character.AI files Chapter 11 bankruptcy, sells assets to Disney for $20M. Character.AI filed for US Chapter 11 bankruptcy and agreed to sell some assets to Disney for $20 million while continuing operations during the sale process. (reuters.com via Techmeme)
• 8. Google's Gemini 4 Argon ties GPT-6 Astra on Artificial Analysis index. Gemini 4 Argon tied GPT-6 Astra at 53 on Artificial Analysis (The Batch), with a launch discount of $1.99/task (60% of Astra's cost) but standard pricing ($3.98) making it pricier than Astra (~$3.32). (Fri Oct 9) (The Batch (DeepLearning.AI), The Batch (DeepLearning.AI))
• 9. Google DeepMind's AlphaProof Nexus solves math problems with peer review. Google DeepMind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems two of which were unsolved for 56 years, per a study in Science. (the-decoder.com via Mark Riedl)
• 10. Anthropic claims small AI models may self-replicate on IoT devices. In Anthropic's simplified evaluations, GLM-5.2 and stronger models are already capable of end-to-end replication. Models of 30B and under complete many individual steps but make mistakes when chaining them. At historical improvement rates, small models match this level by year-end. Critical pass: overstated. (Anthropic)

In brief

• Nvidia AI chip shipments in China drop 44% YoY (reuters.com via Techmeme)
• Humanity's Sixth Sense benchmark reveals AI visual reasoning gaps (Jack Clark)
• Researchers Designed HYSET for LLM Agent Tool Selection (The Batch (DeepLearning.AI))

Open questions

• Will a documented source confirm or refute OpenAI's reported $20B revenue-forecast cut and 2027 IPO timing? Shapes pricing pressure from frontier labs and the exit window for AI startups. Settled by: Reporting that cites investor documents, or an OpenAI statement on forecasts or IPO timing. 1
• Does GPT-6.1 Sol's benchmark parity with Astra hold on real production tasks in portfolio workloads? Decides whether a cost cut of more than 75% becomes real margin. Settled by: Side-by-side evals on portfolio companies' own task sets over the next two weeks. 3
• Will Anthropic publish the method behind its 90–99% claim, and do the defenses degrade output for legitimate users? An older defense scored 71.8% vs 80.4% on SWE-bench Verified; quality cost matters. Settled by: A published method, outside replication, or developer reports on output quality for Fable 5.1 and Opus 5.5. 5

34 items analyzed from 13 sources · 38 filtered as off-topic · 18 more stories in the dashboard. Analysis: claude_code:default, local:qwen3:30b. Our read, the other views and the day in brief are machine-written. Generated 2026-10-13 16:45 EDT.