TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get smart everyday buys delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Five frontier AI models released in September 2026 — Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna — have been compared on the same independent yardstick, the Artificial Analysis Intelligence Index (v4.3). Claude Opus 5.5 posts the top score at 58, while GPT-6 Luna is by far the cheapest at roughly $0.07 per task. The comparison argues that cost per task, not raw intelligence ranking, is the deciding metric for buyers.

Five frontier AI models released within three weeks — Anthropic’s Claude Fable 5.1 and Claude Opus 5.5, and OpenAI’s GPT-6 Astra, GPT-6 Sol and GPT-6 Luna — have been compared on a single independent yardstick, and the results upend the launch-week rankings. On the current Artificial Analysis Intelligence Index (v4.3), Claude Opus 5.5 posts the highest score ever measured by the tracker at 58, while OpenAI’s budget-tier GPT-6 Luna completes indexed tasks for roughly $0.07 each — a spread of nearly 100x in cost per task between the cheapest and most capable models. The comparison’s central finding: ranking models by intelligence alone is the wrong question for buyers; the useful measure is which model clears a quality bar at the lowest cost per task.

The September release calendar was unusually dense. Anthropic shipped Claude Fable 5.1 on September 1 and Claude Opus 5.5 on September 22. OpenAI released GPT-6 Astra on September 3, then GPT-6 Sol and GPT-6 Luna on the same day as Opus 5.5. All five are now measured on the same independent index, allowing a direct score-versus-cost comparison at matched effort levels.

At maximum effort, the results break into three tiers. Opus 5.5 leads on score at 58, with per-token pricing of $4 input / $20 output per million tokens and a cost per indexed task of $5.98. GPT-6 Astra and Fable 5.1 tie at roughly 53, but at very different economics: Astra costs $3.26 per task at $10/$50 pricing, while Fable 5.1 costs $7.63 — now outscored by its cheaper Anthropic sibling. GPT-6 Sol (score 48, $1.06 per task) and GPT-6 Luna (score 37, $0.07 per task) occupy the budget end of the frontier.

The per-task cost picture sharpens the differences. For $100 at maximum effort, the comparison estimates roughly 17 tasks for Opus 5.5, 13 for Fable 5.1, 31 for Astra, 94 for Sol and 1,429 for Luna. Notably, Opus 5.5’s high per-task cost despite mid-tier token pricing is driven by verbosity: at max effort it writes roughly 119,000 output tokens per task, versus about 27,000 for GPT-6 Astra, according to the comparison.

Why Cost per Task Beats Intelligence Rankings

The comparison matters because buyers face a practical decision, not a leaderboard. A team running high-volume classification, extraction or routing can use GPT-6 Luna at sub-cent-to-seven-cent task costs — roughly two orders of magnitude cheaper than flagship models — if its score of 37 meets their quality bar. Teams producing client-ready knowledge work or long agentic coding get the strongest independent score from Opus 5.5, which posted 1822 Elo on the private AA-Briefcase knowledge-work evaluation, 143 points ahead of Fable 5.1 and, per the comparison, the first Anthropic model to beat GPT-5.6 Sol on presentation quality.

The value positioning is also shifting. GPT-6 Astra ties Fable 5.1 on score at well under half the cost per task, making it, in the comparison’s framing, the best value near the top of the table. Fable 5.1 emerges as the model to question: it is now outscored by a cheaper sibling at a higher cost per task, though Anthropic positions it with safeguards suited to workloads where buyers’ own tests show it ahead.

The comparison also flags model-specific trade-offs: most cybersecurity tasks on Anthropic models re-route via fallback to older Claude models, which can depress scores on security and biology work; GPT-6 Sol regressed on polished, complete deliverables (GDPval-AA); and GPT-6 Luna’s Coding Agent Index slipped 2 points versus its predecessor.

The Re-Based Index That Changed the Scores

The most consequential background detail is that Artificial Analysis rebuilt its index this month. When Fable 5.1 launched on September 1, it scored 66 on the older v4.1.1 version of the index — the highest score ever recorded at the time. On the current v4.3 version, the same model scores 53, tied with GPT-6 Astra. The model did not get worse; the test got harder. Any comparison mixing launch-week headlines from different weeks is, as the analysis puts it, comparing different rulers. All figures cited here come from the current v4.3 index.

Two further measurement caveats apply. First, Anthropic models run with fallback behavior: safety-flagged requests route to older Claude models, which can lower scores on security and biology tasks. Second, maximum effort is not the default — most real deployments run at medium or high effort, where rankings and per-task costs can differ from the max-effort figures shown. Fable 5.1’s context window was not listed in the sources used for the comparison, while Opus 5.5 supports 1M tokens, Astra 1.05M, Luna 1M and Sol 872k.

“Ranking them by intelligence is the wrong question. The useful one is which one clears your quality bar at the lowest cost per task.”

— The comparison (Thorsten Meyer AI)

Caveats Buyers Should Weigh

Several limits remain. Fable 5.1’s context window is unlisted in the sources used, leaving an incomplete spec sheet. The index was re-based mid-month, so scores published before the v4.3 change — including Fable 5.1’s launch-week 66 — are not comparable with current figures. Fable 5.1 was measured only at xhigh and max effort, and Sol and Luna only at low and max, so the cost-versus-score curves for intermediate effort levels are partially estimated.

It is also unclear how the Anthropic fallback mechanism affects real-world performance on security and biology tasks beyond its measured effect on index scores, and whether max-effort verbosity (such as Opus 5.5’s ~119k output tokens per task) persists at the medium and high effort levels most deployments actually use. Benchmark scores are historical measurements on specific evaluations, not guarantees of performance on any buyer’s workload.

Testing on Your Own Workload

The comparison’s recommended next step for buyers is a shadow test: run each candidate model on a sample of real tasks from your own workload and compare quality and cost directly, since index results may not transfer. For teams considering Opus 5.5, the guidance is to run at medium effort first and escalate only when needed, given the token volume at max effort. Buyers weighing Astra should evaluate it for computer-use and agent workloads where its token efficiency pays off; those weighing Fable 5.1 should verify with their own tests whether it outperforms Opus 5.5 on their tasks before paying the premium. Further index revisions from Artificial Analysis — which re-based the scale this month — could again shift the numbers, so rankings should be re-checked against the current version before any procurement decision.

Key Questions

Which of the five models scored highest?

Claude Opus 5.5 scored 58 on the Artificial Analysis Intelligence Index (v4.3) at maximum effort — the highest score the tracker has measured, by several points, on the current version of the index.

Which model is the cheapest to run?

GPT-6 Luna, priced at $0.10 input / $0.50 output per million tokens, completes an indexed task for roughly $0.07. For $100 at max effort, the comparison estimates about 1,429 tasks, versus 17 for Opus 5.5.

Why did Claude Fable 5.1’s score drop from 66 to 53?

Artificial Analysis re-based its index from v4.1.1 to v4.3 during September 2026. The same model scored 66 on the old version and 53 on the new one — the model did not change, the test became harder. Scores should only be compared within a single index version.

Why does Opus 5.5 cost more per task than its token price suggests?

At maximum effort, Opus 5.5 writes roughly 119,000 output tokens per task, versus about 27,000 for GPT-6 Astra. Despite less than half of Astra’s per-token price, that verbosity pushes its cost per task to $5.98 — nearly twice Astra’s $3.26.

Should I pick a model based on these benchmark scores alone?

No. The comparison advises running a shadow test on your own tasks before switching. Most deployments run at medium or high effort rather than max, where rankings and costs differ, and Anthropic’s fallback routing to older models can affect results on certain task types.

Source: Thorsten Meyer AI