intelligence-radar/frontier/ledger-2026-09.md

Imported exploration, 2026-09-13. Source claims have not been independently verified by this repository. See integration review.

Frontier ledger — September 2026

Last human pass: 13 September 2026. Prices and ranks rot quickly. Re-verify against Artificial Analysis, lab pricing pages, and Epoch before any purchase.

Figures below are indicative public list prices in USD per million tokens and recent public board readings. Effort settings change everything.

Usable cluster (hosted)

Model	Input / output $/M (list, public reports)	Role in Radar language	Availability note
Claude Fable 5.1	~$10 / $50; cache reads cut to ~$0.25	Frontier-class; often joint AA lead	Generally available; Mythos 5.1 is restricted
GPT-6 Astra	reports vary; treat ~$10 / $50 as working assumption until card is stable	Frontier-class; computer-use / ARC-AGI-3 specialist claims	Rollout gated; Pro capacity incidents mid-Sep
Claude Opus 5	~$5 / $25	Topnotch generalist; often Efficient vs Fable	Broad
GPT-5.6 Sol	promotional cuts reported (~$4 / $20 in some windows)	Topnotch; previous flagship	Broad
GPT-5.6 Terra	lower than Sol (reports ~$2.5 / $12–15)	Often Efficient inside OpenAI family	Broad
GPT-5.6 Luna	lower still	Sufficient / Budget for many knowledge tasks	Broad
Gemini 3.8 Flash	reports ~$0.75 blended class	Sufficient multimodal workhorse; sometimes Budget	GA 2 Sep; watch 2027 price-change notes
Grok 4.6	reports ~$2 / $6	Value hosted frontier-adjacent; long-horizon agents	Current xAI flagship
Muse Spark 1.3	very low list vs US peers on some trackers	Watch for agentic coding efficiency	Closed Meta line
Claude Sonnet 5	cheaper Claude tier	Sufficient for many writing/coding tasks	Broad

Open / exportable cluster

Model	Role	Note
Kimi K3	often strongest open-weight intelligence reading	License and endpoint availability vary by region
Qwen3.8-Max	multilingual + strong general	Open-weight flagship path
DeepSeek V4-Pro / V4.1 Flash	Budget-to-Sufficient coding & general	V4.1 Flash: smaller active params, MIT, 1M context
GLM-5.3 / Flash	open-weight efficiency	Frequent Pareto mention on cost/intelligence
Mistral Medium 3.5	EU-sovereign option	Not always peak score

Composite boards (do not mix versions)

• AA Intelligence Index v4.3 (10 evals including Terminal-Bench 4.0 + AutomationBench-AA): Fable 5.1 max ≈ Astra max ≈ 53; Opus 5 max ≈ 51; Fable 5 ≈ 50; Muse Spark 1.3 max ≈ 48; Sol max ≈ 47. Cost per AA task reported far lower for Astra than Fable 5.1 in AA’s own methodology (~$3.26 vs ~$7.63 in one published comparison).
• Epoch ECI: GPT-6 Astra reported at the top of the current staircase (~169), then Fable 5 / Opus 5 cluster.
• Coding / computer use: SWE-bench Pro, DeepSWE, Terminal-Bench 4.0, OSWorld 2.0, Frontier Code 1.1 — Astra, Fable 5.1, Opus 5, SWE-2, Grok 4.6 trade blows depending on harness.

Radar labels (how the Task Router will speak)

• Frontier — best expected success on the task’s five-benchmark blend, ignoring price. If two models are within noise, name both and pick the one with better availability.
• Topnotch — expected success within 10% relative of Frontier on that blend.
• Efficient — lowest expected task cost among Topnotch (tokens × list price × typical traces, not list price alone).
• Sufficient — judged able to complete the task to an acceptable bar (see METHOD).
• Budget — lowest expected task cost among Sufficient.

Changelog

• 2026-09-13: v0.1 ledger from public boards. Not an audit.