Breaking

OpenAI GPT-6 Astra and Anthropic Claude Fable 5.1 split the frontier on cost, coding, and benchmarks

September 4, 2026 at 16:59 EDT

Foundation Models
Software Dev & Coding
AI Agents

OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1, released two days apart in early September 2026, match on list price but diverge on independent scores, vendor evals, cache economics, and real-world coding and agent work. There is no single winner across the published numbers.

September 2026 · OpenAI · Anthropic

Same list price, split frontier

GPT-6 Astra and Claude Fable 5.1 both bill $10 / $50 per million tokens. Independent indices favor Fable 5.1; several hard vendor evals favor Astra. Cache reads and token use decide real cost.

Cache read price — Fable 5.1 is 4× cheaper on reused context ($0.25 vs $1 per million).

$0.25

Fable 5.1

$1.00

GPT-6 Astra

Four stacked blocks vs one: cache-heavy agent loops still favor Fable 5.1 even at matched list rates.

66 vs 61

AA Intelligence Index

70 vs 67

AA Coding Agent Index

97.6% vs 87.8%

FrontierMath Tier 4 v2

41.4% vs 31.4%

AutomationBench

Where Fable 5.1 leads

Independent intelligence and coding-agent scores, cheaper cache reads, long-horizon agent coding. Anthropic says typical bills drop ~25% vs Fable 5, up to 45% on agentic work.

Where Astra leads — and the catches

Math, automation, several coding and science-agent vendor scores, ~3× fewer coding tokens vs Sol. List price is 2.5× Sol; prompts over 272K reprice the request; rollout is still phased.

No single champion. Match the model to the job.

Fable 5.1 for independent coding-agent lead and cache-heavy loops · Astra for computer use, math, and token-efficient coding vs Sol · Test on your own tasks.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.