This section is not intended as a complete specification sheet. Each row records something the vendor states in the primary sources used for this edition. Items not publicly documented are listed explicitly rather than inferred. Values are time-sensitive and were checked against primary vendor sources on the date given below each table; they should be rechecked before being relied on.
GPT-6 Astra (OpenAI)
| What OpenAI documents | Value or description |
|---|---|
| Context window | 1,050,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | 30 April 2026 |
| Input modalities | Text and image |
| Reasoning control | Five effort levels: low, medium, high, xhigh, max; changeable mid-conversation with cache preserved |
| Long-running mechanisms | Asynchronous tool calling, mid-turn steering, persisted reasoning, compaction, multi-agent orchestration, subagent delegation |
| Tools in the surrounding platform | Web search, file search, code interpreter, hosted shell, computer use, MCP, tool search |
| Input price | $10 per 1M tokens |
| Output price | $50 per 1M tokens |
| Cached input | $1 per 1M tokens; cache writes $12.50 per 1M |
| Long-context pricing | When input exceeds 272,000 tokens, the entire request uses long-context rates: $20 input, $2 cached input, $25 cache writes and $75 output per 1M tokens |
| Safety classification | Critical cybersecurity capability threshold |
| Safeguards described | Isolation, checkpoint encryption, full-trajectory monitoring, blocking alignment evaluations, robustness training, real-time misalignment monitoring |
| Monitorability finding | Lower chain-of-thought monitorability and higher chain-of-thought controllability than GPT-5.6 Sol in adversarial evaluations; stronger action-only monitorability in aggregate |
| Not documented in sources used | Parameter count, internal architecture, routing design, training-data composition, RL recipe, internal representation of persisted reasoning |
Sources: OpenAI model documentation, model guidance, release post, safety overview and system card. Checked 9 September 2026; specifications and pricing re-verified against developers.openai.com on 9 September 2026.
Claude Fable 5.1 (Anthropic)
| What Anthropic documents | Value or description |
|---|---|
| Relationship to Mythos 5.1 | Same underlying model; Fable is the generally available product with additional safeguards, Mythos is available to vetted organizations through trusted-access programs |
| Context window | 1,000,000 tokens (default and maximum), standard pricing across the whole window |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | June 2026 |
| Input modalities | Text and images; diagrams, charts and tables inside files and PDFs |
| Reasoning control | Adaptive thinking, always on; effort parameter with five levels: low, medium, high, xhigh, max; default high; per-message effort change preserving cache (beta) |
| Long-running behavior | Plans work, uses tools, recovers when a step fails, operates across several applications, continues unattended for long periods; managed agents |
| Self-verification | Writes its own tests; uses vision to compare outputs against a design or goal |
| Input price | $10 per 1M tokens |
| Output price | $50 per 1M tokens |
| Cache writes | $12.50 per 1M (5-minute); $20 per 1M (1-hour) |
| Cache reads | $0.25 per 1M tokens, described as reducing agentic workload cost relative to Fable 5 |
| Batch API | 50% discount on input and output |
| Release | 1 September 2026; retirement not before 1 September 2027 |
| Safeguards described | Additional classifiers for sensitive cyber, biology and chemistry use; routing of some requests to Opus models without the sensitive-capability exposure |
| Research tooling | Claude Science workbench connecting models to scientific databases, notebooks, R, cluster terminals; auditable artifacts |
| Physical-device interface | Model Hardware Standard research preview (27 August 2026, with HHMI Janelia): microscopes, liquid handlers, robotic arms; parallel instrument operation; safety limits enforced at the driver level below the agent; some hardware-error recovery; open-sourcing planned after further safety work |
| Not documented in sources used | Parameter count, internal architecture, routing design inside the model, training-data composition, RL recipe, internal state mechanism for long-running work |
Sources: Anthropic model documentation (platform.claude.com), product pages for Fable and Mythos, safeguards announcement, Claude Science announcement, Model Hardware Standard preview, system-cards index. Checked 9 September 2026; specifications, effort control and pricing re-verified against platform.claude.com on 9 September 2026.
Other figures cited in the book
| Figure | Where it appears | Status |
|---|---|---|
| OpenAI research organization using about 3.1 agent-workdays per human workday by mid-August 2026; more than half of successful 4–8 hour agent tasks still needed human intervention | Chapter 16 | OpenAI self-report, 6 September 2026; re-verified; compute-growth and correlation caveats stated in source |
| Astra completing OSWorld-style tasks faster than GPT-5.6 Sol with a higher score | Chapter 7 | OpenAI release post, 3 September 2026 |
| Astra solving harder mathematics in a single forward pass in one external evaluation | Chapter 14 | OpenAI system card, September 2026 |