GPT-6 Astra Matches Fable 5 on Coding Agents but Costs 75% More Per Task Than Its Own Predecessor
Artificial Analysis published its GPT-6 Astra benchmarks on September 3, and the results split cleanly along workload type. On coding agent tasks, Astra is a strong value proposition. On general intelligence tasks, it costs significantly more than the predecessor it replaces.
Coding Agent Index: The Good News
In Artificial Analysis’s Coding Agent Index — which runs models through agentic code completion tasks in a Codex harness — GPT-6 Astra scores 67. That puts it approximately equal to Claude Opus 5 and Fable 5 run in Claude Code. Claude Fable 5.1 leads the index at 70.
The reason Astra matches Fable 5 at a fraction of the cost is token efficiency. Astra uses roughly one-third the tokens of GPT-5.6 Sol in the same harness, and about one-fifth the tokens of Claude Opus 5 at extended effort. That translates directly to per-task cost:
| Model | Coding Agent Index | Approx. Cost vs Astra |
|---|---|---|
| Claude Fable 5.1 (Claude Code) | 70 | ~2x |
| GPT-6 Astra (Codex, max) | 67 | baseline |
| Claude Fable 5 (Claude Code) | 67 | ~2x |
| Claude Opus 5 (xhigh) | ~67 | ~5x |
| GPT-5.6 Sol (max) | ~65 | ~1x (same cost) |
Astra matches Sol’s per-task cost while scoring 2 points higher on the index. Against Fable 5, it delivers the same score at less than half the price. For coding agent workloads specifically — automated pipelines, CI/CD integration, multi-step repository tasks — Astra appears to be the current price/performance leader on this axis.
Intelligence Index: The Problem
On the Artificial Analysis Intelligence Index, Astra scores 61 — the same as GPT-5.6 Sol, and 5 points below Fable 5.1 (66). It also trails Meta’s Muse Spark 1.3 at max settings.
The score itself is not the issue. The issue is what Astra costs relative to its predecessor:
- GPT-5.6 Sol pricing: $4/M input, $20/M output
- GPT-6 Astra pricing: $10/M input, $50/M output — 2.5x across the board
- Token efficiency gain: ~10% fewer output tokens at max effort
- Net per-task cost at max effort: 75% more than Sol for the same Intelligence Index score
The cache discount structure is the same (90% off cache reads, 25% premium on cache writes). The efficiency gain from reduced token use does not offset the price increase for general knowledge tasks.
On AA-Omniscience, Astra’s hallucination benchmark, there is meaningful improvement: hallucination rate dropped from 92% to 51% at max effort, with accuracy increasing by 4 points simultaneously. That is a significant reliability gain, not a tradeoff. Whether it justifies the price premium depends on whether your use case is sensitive to hallucination rates — long-form research synthesis, document extraction, enterprise RAG pipelines.
On AA-Briefcase, Artificial Analysis’s long-horizon knowledge work evaluation, Astra gains roughly 80 Elo points over Sol. The benchmark runs multi-week simulated projects with linked tasks and thousands of source files. Astra improves on rubric scores and Analytical Quality Elo. GPT-5.6 Sol retains the Presentation Quality Elo lead.
Pricing
GPT-6 Astra — $10/M input | $50/M output
Cache reads: 90% discount ($1/M input cached)
Cache writes: 25% premium ($12.50/M)
Context: 1,050,000 tokens
Max output: 128,000 tokensThe Split Verdict
For agentic coding workloads: Astra is competitive with the frontier at better economics than Fable 5, roughly on par with Sol’s cost at higher coding benchmark performance.
For general intelligence tasks at max effort: Astra costs 75% more than Sol for the same output quality, and less than Sol costs is spent to reach Fable 5.1’s performance level. The value case depends heavily on whether the hallucination improvement and AA-Briefcase gains matter to the workload.
OpenAI is effectively asking customers to pay a 2.5x price premium for a model that, on general tasks, performs where Sol already was — while delivering a meaningful coding agent capability jump and substantially reduced hallucination rates. That is a coherent product decision if the target buyer is agentic coding pipelines, not general API usage.
Whether the hallucination and long-horizon reasoning improvements justify the cost for non-coding workloads is a decision each buyer will have to run for their specific tasks. Artificial Analysis’s methodology runs standardized harnesses; individual results vary by prompt structure, caching patterns, and effort level.
GPT-6 Astra is available now in API. Coding Agent Index benchmark data is available at artificialanalysis.ai.