The Memo - Special edition - Claude Fable 5.2 and Claude Mythos 5.2 - 9/Nov/2026

1. Key Themes

Theme: AI is now building AI, and the most capable models are being withheld from the public

Claude writes most of Claude

Anthropic's core codebase is now predominantly AI-authored, a concrete sign of recursive improvement.

"Anthropic says that over 90% of the changes merged into its core codebase are now written by Claude, with human guidance."

Internal models are far ahead of public releases

The author argues the public frontier has been decoupled from the true frontier.

"Fable and Mythos are no longer frontier models."

"METR reports average self-reported uplift of about 12x. In plain English, Anthropic's own staff say they're getting about a year's worth of work done every month."

Withholding capability is now a leading indicator of ASI

"one of the major leading indicators of ASI is that the most capable models would no longer be publicly available. Models 3 and 4 are a clear example of this indicator."

Theme: Long-horizon agentic work at massive parallel scale

Swarms of subagents complete multi-day engineering projects

"The model coordinated more than 1,000 subagents working around the clock for eight days. The result boots on a laptop, and includes a kernel, disk, keyboard, and network drivers, file storage, windows, a login screen, an installer, and built-in Claude features."

Formal mathematics falls to agent swarms

"1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours. An unreleased research version of the model took 11 days. Opus 5.5 with the same setup finished only 26% of the formalization in 18 hours."

Theme: AI is moving into hardware and the physical world

End-to-end electronics design in under a day

"Working with Diode, a company that designs and manufactures circuit boards, the model designed a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet... All three devices worked without any electrical changes, and each board was ready for manufacturing in under a day."

Robotics generalization with a large benchmark jump

"Anthropic says Fable 5.2's ability to control physical robots is much better than earlier Claude models. That includes robotics tasks it was never explicitly trained on."

"On the RoboDojo Sim benchmark, Fable 5.2 scores 64.9%, more than double Fable 5.1 (27.6%) and GPT-6 Astra (29.8%), and well ahead of Opus 5.5 (39.4%)."

Theme: Frontier AI is accelerating scientific discovery, with tiered access for dangerous domains

Open problems in theoretical physics

"Fable 5.2 solved two open problems in black hole physics. The first completes a calculation started in 2003... The second finds the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity."

Structural biology relevant to gene therapy

"The model accurately modeled 28 of the 34 adeno-associated virus (AAV) capsid structures it attempted... AAV capsids are the shells used to deliver gene therapies."

Dual-tier release model: same weights, different safeguards

"Fable 5.2 and Mythos 5.2 are the same model with two different sets of safeguards. Fable 5.2 has strong cybersecurity and biology safeguards, and is available to everyone. Mythos 5.2 has some of those safeguards removed, and is only available to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs."

Theme: Rapid price/performance compression and benchmark saturation

Capability that was restricted seven months ago is now public and cheaper

"Just seven months ago, Anthropic told us that Mythos Preview was too dangerous to release... Today, the same class of model in Fable 5.2 is available to the public, at less than a third of what Mythos Preview cost the Glasswing partners."

Pricing falls while speed rises

"Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, down from $10/$50 for Fable 5. An optional fast mode generates tokens up to 2.5x faster for $16/$80."

Benchmarks are saturating

"Claude Fable 5.2 scores HLE=74.2% (with tools), well above Opus 5.5 (67.7%) and GPT-6 Astra (57.2%)."

"'We consider GPQA Diamond to be a saturated evaluation and plan to stop reporting the performance of future models on it.'"

2. Contrarian Perspectives

Model size no longer predicts performance, and a price cut doesn't mean a smaller model

The author pushes back on the intuitive reading that cheaper, faster models must be smaller, attributing the gains to inference efficiency.

"General model size is no longer an indicator of performance, but I still find it interesting."

"A price cut alongside a speed increase on the same base points to inference efficiency gains rather than a smaller model."

The publicly available "frontier" is no longer the real frontier

Public benchmark leaderboards understate where capability actually sits, because labs are keeping the strongest systems internal.

"Fable and Mythos are no longer frontier models... Anthropic's internal models don't have that limit, and they have leapt well past anything publicly available."

A lab deliberately slowing its own model on recursive self-improvement

Rather than racing to maximize every capability, Anthropic reportedly capped the skills that most directly accelerate AI R&D. The author sees this as a notable safety-driven precedent.

"Fable 5.2 and Mythos 5.2 perform at about the level of Opus 5.5. Anthropic says this did not affect agentic coding, hardware development, or traditional machine learning research. This is the first time I have seen a frontier lab publicly say it deliberately slowed one of its own models in the area that matters most for recursive self-improvement."

3. Companies Identified

Anthropic

• Description: AI lab developing the Claude model family
• Why mentioned: Released Fable 5.2 and Mythos 5.2, holds back the most capable internal models, and reports 90%+ AI-written code
• Quotes: "Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2." / "Anthropic took steps to limit the research capabilities that most directly accelerate AI development."

Diode

• Description: Company that designs and manufactures circuit boards
• Why mentioned: Partner in the electronics design demonstration
• Quotes: "Working with Diode, a company that designs and manufactures circuit boards, the model designed a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet."

METR

• Description: Organization that assesses AI capabilities and R&D acceleration
• Why mentioned: Provided the assessment showing ~12x uplift at Anthropic from internal models
• Quotes: "METR's latest assessment of AI R&D acceleration at Anthropic (12/Aug/2026 to 29/Oct/2026) covers two unreleased internal models."

OpenAI (GPT-6 Astra)

• Description: Competing frontier lab; GPT-6 Astra is its model
• Why mentioned: Benchmark comparison point and cited as proto-ASI
• Quotes: "As with earlier models including Mythos 5, Fable 5, and GPT-6 Astra, both new models should be considered proto-ASI."

Glasswing partners

• Description: Select group of trusted partners given early access to Mythos Preview
• Why mentioned: Pricing and access comparison for the restricted-release model
• Quotes: "...at less than a third of what Mythos Preview cost the Glasswing partners."

4. People Identified

Dr Alan D. Thompson

• Description: Author of LifeArchitect.ai and The Memo
• Why mentioned: Provides the ASI framework, HLE/GPQA criteria, and the model size estimates
• Quotes: "'It is likely that any model with a primary score at >50% on HLE, and a secondary score at >90% on GPQA is an ASI system.'"

5. Operating Insights

Use hierarchical agent orchestration for large, multi-day builds

The OS prototype and Fermat formalization both relied on very large agent swarms running continuously, suggesting operators should design workflows around parallel subagents rather than single long sessions.

"The model coordinated more than 1,000 subagents working around the clock for eight days."

Compress hardware development cycles with AI-driven design

Teams building physical products can use a one-paragraph prompt to get circuit boards, enclosures, and software, cutting design-to-manufacturing time.

"Each board was ready for manufacturing in under a day."

Budget for falling costs and use fast-mode tiers strategically

Prices fell while speed rose, and a premium fast tier now exists for latency-sensitive uses.

"Priced at $8 per million input tokens and $40 per million output tokens, down from $10/$50 for Fable 5. An optional fast mode generates tokens up to 2.5x faster for $16/$80."

6. Overlooked Insights

Verification programs as a gating mechanism for dangerous capabilities

Access to the less-restricted Mythos 5.2 runs through vetted programs, hinting at an emerging compliance/verification layer that could create business opportunities for organizations in security and life sciences.

"Mythos 5.2 has some of those safeguards removed, and is only available to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs."

Benchmark ceilings are limited by dataset errors

The usefulness of top benchmarks is capped by label quality, which matters for anyone relying on leaderboard scores for vendor selection.

"Based on the highest testing suites we have, and the estimated ceiling for HLE due to errors (HLE≈51.3%), this model continues to meet the HLE half of my criteria for an early ASI system."