My September 2026 AI Stack By Role And Responsibility

🔍 Read the full analysis: My September 2026 AI Stack By Role And Responsibility on ThorstenMeyerAI.com

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

Get office and shipping supplies 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

Thorsten Meyer published his working AI stack for September 2026 on 29 September 2026, assigning roles to six frontier models as the market shifted from a capability race to a price contest. His setup: Claude Opus 5.5 at high or extra-high effort as the main builder, the just-released GPT-6.1 Sol for detail work and review, and the ultra-cheap GPT-6 Luna for routing decisions. The recommendation matters because six leading models now sit within about 20 points on the Artificial Analysis Intelligence Index while their cost per task differs by roughly 100x, turning model choice into an economics question rather than a leaderboard question.

Meyer’s core finding is that the question has changed from “which model is smartest?” to “which model clears my quality bar at the lowest cost per task?” According to his analysis, based on the Artificial Analysis Intelligence Index v4.3.x, Opus 5.5 (released 22 September) posts the highest score at 58 on its max setting, at $5.98 per task — about 17 tasks per $100. GPT-6.1 Sol at extra-high effort scores 51 for $0.39 per task, or 256 tasks per $100, while GPT-6 Luna scores 37 for $0.07 per task, or 1,429 tasks per $100.

Three cost anomalies stand out in his data. Opus 5.5 now outscores its more expensive sibling Fable 5.1 by 5 points while costing less per task. Sonnet 5.5 at max effort costs more per task than Opus at max while scoring 2 points lower. And GPT-6.1 Sol costs roughly one-eighth of Astra and one-twentieth of Fable per task for a score only 1 to 2 points lower. Per-token pricing cited: Opus 5.5 at $4 input / $20 output per million tokens, Fable and Astra at $10 / $50, Sol at $2 / $10, and Luna at $0.10 / $0.50.

Meyer also argues that the effort setting is the biggest cost lever. On Opus 5.5, moving from extra-high to max adds 2 index points for 73% more cost per task; moving from medium to max raises cost 4.46x for 7 points. Sonnet 5.5 at max writes about 193k output tokens per task — the most Artificial Analysis has measured — pushing its cost from $2.74 to $7.60 for 4 points. GPT-6.1 Sol has real tradeoffs too: its high and extra-high settings take 57 to 69 seconds to produce a first token, making it unsuitable for interactive use, and Opus still leads it by 5 points at extra-high effort.

Why Cost Per Task Now Drives Model Choice

The stack reflects a structural change in how AI tools are selected. When top models cluster within roughly 20 index points but differ by two orders of magnitude in cost, the dominant variable for most workloads becomes price per completed task, not peak benchmark score. For developers and teams, this means effort settings and model-role assignments — not model purchases — are the main budget controls.

The review role is the practical centerpiece. Meyer argues that a different model family reviewing Opus output is a stronger check than Opus reviewing itself, and at $0.32 to $0.39 per task, review passes become cheap enough to run on every meaningful change. He also cautions that cheaper tokens are not cheaper work: in an illustrative (not measured) example, halving model price saves only 12.5% of real cost, and a single extra minute of human review erases the saving.

A Month of Six Frontier Releases

September 2026 brought a rapid succession of frontier releases, according to Meyer’s timeline: Claude Fable 5.1 on 1 September, GPT-6 Astra on 3 September, GPT-6 Luna and Claude Opus 5.5 on 22 September, Claude Sonnet 5.5 on 28 September, and GPT-6.1 Sol on 29 September — the day of publication. Sol launched at the same $2 / $10 per-million-token pricing as its week-old predecessor; even its medium setting matches the earlier GPT-6 Sol’s score of 48 at one-fifth of that model’s $1.06 per task, per Artificial Analysis figures. Sol is also unusually concise, using 25M output tokens on the index at its high setting against a median of 82M for comparable models.

All scores in the analysis come from the Artificial Analysis Intelligence Index v4.3.x, which Meyer describes as a map of general capability, not a verdict on any specific workload — his standing advice is to shadow-test before switching models.

“In four weeks, the AI frontier stopped being a leaderboard and became a price curve.”

— Thorsten Meyer

Shadow Tests and the Next Release Cycle

Meyer’s stated practice is to shadow-test any candidate model against real workloads before making it a default, and his stack is dated explicitly to 29 September 2026 — a snapshot he expects to revise. The immediate open items are Artificial Analysis’s publication of Sol’s low and max effort settings, and whether the October release cycle compresses or widens the price-performance gap. Readers applying the stack to their own work should treat the effort-level tables as starting points and validate against their own task distributions and quality bars.

Key Questions

Which model does Meyer recommend as the main builder?

Claude Opus 5.5 at high effort — 54 index points for $1.82 per task — with extra-high effort (56 points, $3.46) reserved for hard problems such as architecture, migrations and trust boundaries. Max effort is described as rarely worth the cost.

Why use GPT-6.1 Sol instead of a higher-scoring model?

Sol costs about $0.39 per task at extra-high effort versus $3.26 for Astra or $7.63 for Fable, while scoring only 1 to 2 points lower. That makes routine review passes and detail work affordable, though high and extra-high settings have 57-to-69-second times to first token, so it is not an interactive model.

What does GPT-6 Luna do in the stack?

Luna, at 37 index points and $0.07 per task, handles classification, extraction and routing — high-volume work where capability demands are low and cost dominates.

Is the effort setting really more important than model choice?

Per Meyer’s data, often yes: on Opus 5.5, going from medium to max effort raises cost 4.46x for 7 index points. He recommends medium as the default for documents and everyday work, and high as Sonnet 5.5’s best value at 47 points for $1.08.

Are these benchmark scores a guarantee of performance on my tasks?

No. All scores come from the Artificial Analysis Intelligence Index v4.3.x, which measures general capability. Meyer’s standing advice is to shadow-test any model against your own workload before switching, and he notes that one index point is within measurement noise.

Source: ThorstenMeyerAI.com