TL;DR
Play games on Amazon Luna for your game nights, included with Prime
- A rotating selection of games, no download needed
- Play on TV, laptop or phone
- Fast, free delivery for your gear, too
An analysis of the five frontier models released in September 2026 — Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna — finds Opus 5.5 posts the highest score ever measured on the Artificial Analysis Intelligence Index (58), while GPT-6 Luna delivers tasks at roughly $0.07 each. The comparison argues cost per task, not raw intelligence ranking, is the decisive metric for buyers.
Five frontier AI models released within three weeks in September 2026 — Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 and OpenAI’s GPT-6 Astra, GPT-6 Sol and GPT-6 Luna — have been compared on a single independent yardstick, with Claude Opus 5.5 scoring 58, the highest result the Artificial Analysis Intelligence Index has recorded, while GPT-6 Luna completes index tasks for about $0.07. The analysis argues that the decisive question for buyers is not which model is smartest, but which one clears a given quality bar at the lowest cost per task — a figure that varies by a factor of roughly 85 across the five releases.
Anthropic released Claude Fable 5.1 on September 1 and Claude Opus 5.5 on September 22. OpenAI released GPT-6 Astra on September 3, followed by GPT-6 Sol and GPT-6 Luna on September 22. All five were measured on the current Artificial Analysis Intelligence Index v4.3.
At maximum effort, the scores and costs per task are: Opus 5.5 at 58 for $5.98, Fable 5.1 at 53 for $7.63, GPT-6 Astra at 52.7 for $3.26, GPT-6 Sol at 48 for $1.06, and GPT-6 Luna at 37 for $0.07. Pricing per million tokens ranges from Luna’s $0.10 input / $0.50 output to Fable 5.1 and Astra at $10 / $50. Expressed as tasks per $100, the spread runs from 17 for Opus 5.5 and 13 for Fable 5.1 to 31 for Astra, 94 for Sol and 1,429 for Luna.
Each model lands in a distinct niche, according to the analysis. Opus 5.5 is positioned for client-ready knowledge work and long agentic coding, with the strongest showing on the private AA-Briefcase knowledge-work evaluation, where it scored 1,822 Elo — 143 points ahead of Fable 5.1 — and became the first Anthropic model to beat GPT-5.6 Sol on presentation quality. Astra suits computer use and agents where token efficiency matters. Sol is described as a budget workhorse for coding agents and business automation, though it regressed on polished, complete deliverables in the GDPval-AA evaluation. Luna fits high-volume classification, extraction and routing, though its Coding Agent Index slipped 2 points versus its predecessor.
How Do Five Frontier AI Models Measure Up?
Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna — five frontier models released within three weeks — compared on a single independent yardstick. The verdict: cost per task, not raw intelligence, decides the winner.
Score vs. Cost Per Task at Maximum Effort
All five models were measured on the Artificial Analysis Intelligence Index v4.3 at maximum effort. Fable 5.1 ties Astra on score — at more than double the cost per task.
| Model | Vendor | Released | Index Score | Cost / Task | Tasks / $100 | Price / M Tokens (In / Out) |
|---|---|---|---|---|---|---|
| Opus 5.5 | Anthropic | Sept 22 | 58 | $5.98 | 17 | $5 / $25 |
| Fable 5.1 | Anthropic | Sept 1 | 53 | $7.63 | 13 | $10 / $50 |
| GPT-6 Astra | OpenAI | Sept 3 | 52.7 | $3.26 | 31 | $10 / $50 |
| GPT-6 Sol | OpenAI | Sept 22 | 48 | $1.06 | 94 | $1 / $5 |
| GPT-6 Luna | OpenAI | Sept 22 | 37 | $0.07 | 1,429 | $0.10 / $0.50 |
What $100 Buys: Tasks per Hundred Dollars
The same $100 buys anywhere from 13 to 1,429 index tasks depending on which model is selected. Model choice is now a workload decision, not a ranking decision.
Five Models, Five Distinct Jobs
Each model lands in a distinct niche, according to the analysis. Buyers should match workload to model, not chase the top of the leaderboard.
Claude Opus 5.5
Client-ready knowledge work and long agentic coding. Strongest on the private AA-Briefcase evaluation at 1,822 Elo — 143 points ahead of Fable 5.1 — and the first Anthropic model to beat GPT-5.6 Sol on presentation quality.
PICK: HIGHEST QUALITYGPT-6 Astra
Computer use and agents where token efficiency matters: roughly 27,000 output tokens per task versus Opus 5.5’s 119,000. Ties Fable 5.1 on score for well under half the cost per task.
PICK: AGENTS & COMPUTER USEGPT-6 Sol
A budget workhorse for coding agents and business automation — though it regressed on polished, complete deliverables in the GDPval-AA evaluation. 94 tasks per $100.
PICK: BUDGET WORKHORSEGPT-6 Luna
High-volume classification, extraction and routing at sub-cent task costs — viable at a scale no top-tier model can match. Coding Agent Index slipped 2 points versus its predecessor.
PICK: HIGH-VOLUME PIPELINESClaude Fable 5.1
The model buyers should question most carefully: outscored by its cheaper sibling Opus 5.5 at a lower price, and tied on score by Astra at well under half the cost per task. Highest cost per task of the five.
QUESTION THE VALUEShadow Test First
Run shadow tests on your own tasks before switching models. Most production traffic runs at medium or high effort — not max — where rankings and costs differ from headline tables.
TEST, DON’T TRUSTCheap Tokens ≠ Cheap Tasks
Claude Opus 5.5 — Max Effort
GPT-6 Astra — Max Effort
Price per token is a poor predictor of price per task. Opus 5.5’s per-token price is less than half of Astra’s, yet at maximum effort it costs nearly twice as much per task — because it writes far more output. At the budget end, Luna’s sub-cent tasks make volume classification and routing viable at unmatched scale.
Fable 5.1 Didn’t Get Worse — The Test Got Harder
Artificial Analysis rebuilt its Intelligence Index in September 2026, moving from v4.1.1 to v4.3. Launch-week headlines are now incomparable with current numbers. The model didn’t change — the yardstick did.
Sept 1 · Fable 5.1 Launch
Scores 66 on the older v4.1.1 index — described at the time as the highest score ever recorded.
Sept 2026 · Index Rebase
Artificial Analysis rebuilds the Intelligence Index from v4.1.1 to v4.3, making the test harder across the board.
Same Model, New Score
Fable 5.1 scores 53 on v4.3 — tied with GPT-6 Astra. The model itself is unchanged.
Buyer Rule
Only compare scores from the same index version — currently v4.3. Future re-bases could shift scores again.
What the Headlines Leave Out
Several limitations shape these figures — and the buyer’s due-diligence checklist.
Max Isn’t the Default
Most real deployments run at medium or high effort, where both rankings and costs differ from the max-effort table. Fable 5.1 was plotted at xhigh and max only; Sol and Luna at low and max.
Anthropic Re-Routing
Anthropic models run with fallback: safety-flagged requests route to older Claude models. Most cybersecurity tasks for Opus 5.5 re-route to Opus 4.8, so published security and biology scores may not reflect the newest model’s own capabilities.
Gaps in the Curve
Fable 5.1’s context window isn’t listed in the sources; Opus 5.5’s output speed was measured only at medium and xhigh effort. The full effort-versus-cost curve isn’t available for every model.
Quick Answers for Buyers
Which of the five models scored highest?
Claude Opus 5.5, at 58 on the Intelligence Index v4.3 at max effort — the highest score the index has ever measured, by several points.
Why did Fable 5.1’s score drop from 66 to 53?
It didn’t drop — the index was re-based from v4.1.1 to v4.3, making the test harder. Only scores from the same index version should be compared.
Which model is cheapest to run?
GPT-6 Luna, at roughly $0.07 per index task at max effort — about 1,429 tasks per $100, at $0.10 input / $0.50 output per million tokens.
Why is Opus 5.5 pricier per task than Astra despite cheaper tokens?
At max effort Opus 5.5 writes about 119,000 output tokens per task versus roughly 27,000 for Astra — nearly doubling total cost per task despite lower per-token pricing.
What should buyers do before choosing?
Run a shadow test on your own workload at the effort level you actually use — most deployments run at medium or high effort, not max, where rankings and costs differ.
What open questions remain?
Whether Anthropic responds to Fable 5.1’s position as the outscored, higher-priced sibling, whether Sol fixes its GDPval-AA regression on complete deliverables, and how the cost frontier shifts as pricing and effort tuning evolve.
Why Cost Per Task Decides the Winner
The comparison shows that price per token is a poor predictor of price per task. Opus 5.5’s per-token price is less than half of GPT-6 Astra’s, yet at maximum effort it costs nearly twice as much per task — about $5.98 versus $3.26 — because it writes far more output: roughly 119,000 output tokens per task compared with about 27,000 for Astra, according to the analysis.
The gap matters most at the budget end. For workloads that tolerate lower scores, GPT-6 Luna’s sub-cent tasks make volume classification and routing viable at a scale no top-tier model can match. Meanwhile, Astra ties Fable 5.1 on score for well under half the cost per task, and Opus 5.5’s cheaper sibling outscored Fable at a lower price — making Fable 5.1 the model buyers should question most carefully.
For businesses, the practical takeaway is that model choice is now a workload decision, not a ranking decision: the same $100 buys anywhere from 13 to 1,429 index tasks depending on which model is selected.
As an affiliate, we earn on qualifying purchases.
The Index Rebase That Broke Launch Comparisons
Artificial Analysis rebuilt its Intelligence Index in September 2026, moving from v4.1.1 to v4.3, and the change makes launch-week headlines incomparable with current numbers. When Fable 5.1 launched on September 1 it scored 66 on the older v4.1.1 index, described at the time as the highest ever recorded. On the current v4.3 version, the same model scores 53, tied with GPT-6 Astra. The model did not change — the test got harder.
Two further caveats shape the numbers. Anthropic models in the test run with fallback: safety-flagged requests route to older Claude models, which can lower scores on security and biology work — the analysis notes that most cybersecurity tasks for Opus 5.5 re-route to Opus 4.8. And maximum effort is not the default; most real deployments run at medium or high effort, where both rankings and costs differ from the max-effort table. Anthropic positions Opus 5.5 as matching Fable 5.1 on most work at 40% less cost than Opus 5; the independent index places it clearly ahead instead.
Caveats in the Numbers
Several limitations remain. Fable 5.1’s context window is not listed in the sources used for the comparison, and Opus 5.5’s output speed was measured only at medium and xhigh effort. Fable 5.1 was plotted at xhigh and max effort only, and Sol and Luna at low and max, so the full effort-versus-cost curve is not available for every model.
The index itself is a moving target: it was re-based this month, and future re-bases could again shift scores. Anthropic’s safety fallback routing to older models means published security and biology scores may not reflect the newest model’s own capabilities. Medium- and high-effort rankings — where most production traffic runs — are not fully reported here, and results on any individual buyer’s workload may diverge from index-level findings.
Shadow Tests Before Switching Providers
The analysis recommends that organizations run shadow tests on their own tasks before switching models, since index rankings at max effort may not hold at the medium or high effort levels most deployments use. Buyers comparing launch-week coverage from different weeks should confirm all figures come from the same index version — currently v4.3.
Open questions include whether Anthropic responds to Fable 5.1’s position as the outscored, higher-priced sibling, whether OpenAI’s Sol addresses its regression on complete deliverables in GDPval-AA, and how the cost frontier shifts as providers adjust pricing or effort tuning in the coming months.
Key Questions
Which of the five September 2026 models scored highest?
Claude Opus 5.5, at 58 on the Artificial Analysis Intelligence Index v4.3 at max effort — described as the highest score the index has measured, by several points.
Why did Claude Fable 5.1’s score drop from 66 to 53?
It didn’t drop — the index was re-based from v4.1.1 to v4.3, making the test harder. The same model scores 66 on the old version and 53 on the current one. Only scores from the same index version should be compared.
Which model is cheapest to run?
GPT-6 Luna, at roughly $0.07 per index task at max effort, with pricing of $0.10 input / $0.50 output per million tokens. That yields about 1,429 tasks per $100.
Why does Opus 5.5 cost more per task than GPT-6 Astra despite cheaper tokens?
At max effort Opus 5.5 writes about 119,000 output tokens per task versus roughly 27,000 for Astra, so its total cost per task (~$5.98) is nearly double Astra’s (~$3.26) despite lower per-token pricing.
What should buyers do before choosing a model?
Run a shadow test on your own workload at the effort level you actually use, since most deployments run at medium or high effort rather than max, where rankings and costs differ.
Source: Thorsten Meyer AI
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
