My AI Stack For September 2026: Tools And Tasks

🔍 Read the full analysis: My AI Stack For September 2026: Tools And Tasks on ThorstenMeyerAI.com

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

Get everyday essentials 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 AI tool stack for September 2026 on 29 September 2026, naming Claude Opus 5.5 at high or extra-high effort as his main model for building software and the same-day-launched GPT-6.1 Sol as his model for detail work and review. His core finding: six frontier models now sit within about 20 index points of each other on the Artificial Analysis Intelligence Index v4.3.x, while their cost per task differs by roughly 100 times — shifting the practical question from which model is smartest to which one clears a quality bar at the lowest cost per task.

According to Meyer’s analysis, Opus 5.5 (released 22 September) tops the index at 58 points at its max setting, costing $5.98 per task and delivering about 17 tasks per $100. At the other end, GPT-6 Luna scores 37 but costs $0.07 per task, yielding 1,429 tasks per $100. Between them sit Sonnet 5.5 (56 points, $7.60 per task), Fable 5.1 (53, $7.63), GPT-6 Astra (53, $3.26) and GPT-6.1 Sol at xhigh (51, $0.39).

Three findings stand out in his comparison. Opus 5.5 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 for 2 fewer points, which Meyer argues makes it hard to justify at that setting. 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. All scores come from the Artificial Analysis Intelligence Index v4.3.x, which Meyer cautions is a map of general capability, not a verdict on any specific workload.

Meyer also reports that the effort setting is the largest cost lever. On Opus 5.5, moving from xhigh to max adds 2 index points and 73% more cost per task; from medium to max, cost rises 4.46 times for 7 points. GPT-6.1 Sol, launched 29 September at $2/$10 per million tokens, matches its predecessor’s score at medium (48) for one-fifth of the cost per task, though its high and xhigh settings take 57 to 69 seconds to produce a first token, ruling out interactive use at those levels. It is also very concise: 25 million output tokens on the index at high, versus a median of 82 million for comparable models.

Why Cost Per Task Now Drives Model Choice

The significance of the piece is its framing: when frontier scores cluster tightly, the differentiator becomes price and workflow fit rather than raw capability. Meyer’s stack treats models as roles — builder, reviewer, router, bulk classifier — instead of a single default. The review seat is the most consequential: a different model family checking Opus’s output is, in his view, a better safeguard than Opus reviewing itself, and at $0.39 per task it can run on every meaningful change.

He also pushes back on price-focused optimization. Halving model price, he writes, saves only about 12.5% of real cost in his illustrative example, and a single extra minute of human review can erase that saving — meaning cheaper tokens do not automatically mean cheaper work. A small routing model, Jev, which he notes cannot write a sentence, now handles high-volume yes/no and routing decisions in his setup.

A Month of Closely Spaced Releases

The September 2026 stack reflects an unusually crowded release window. Fable 5.1 arrived 1 September, GPT-6 Astra on 3 September, Opus 5.5 and GPT-6 Luna on 22 September, Sonnet 5.5 on 28 September, and GPT-6.1 Sol on 29 September — the same day Meyer published. GPT-6.1 Sol launched at the same per-token pricing ($2 input / $10 output per million tokens) as its week-old predecessor, with Artificial Analysis already listing three effort levels for it.

Per-million-token prices, according to Meyer’s data: Opus 5.5 at $4/$20 with cache reads at $0.20; Fable and Astra at $10/$50; GPT-6.1 Sol at $2/$10; and Luna at $0.10/$0.50. Meyer publishes this stack monthly, and the previous month’s framing, he writes, still treated the frontier as a leaderboard rather than a price curve.

“In four weeks, the AI frontier stopped being a leaderboard and became a price curve. Six models now sit within about 20 index points of each other, while their cost per task differs by roughly 100×.”

— Thorsten Meyer

Benchmarks, Noise, and Unpublished Settings

Meyer explicitly hedges several findings. The Artificial Analysis index is a measure of general capability, not of any particular workload, and he advises shadow-testing before switching models. He notes that one index point is inside the noise, which affects comparisons between closely scored models such as Sol xhigh (51), Astra (52–53) and Fable (53).

Artificial Analysis has not yet published low or max effort settings for GPT-6.1 Sol, so its full cost-performance range is unknown. His cost-savings example — that halving model price saves 12.5% of real cost — is described as illustrative, not measured. Whether Opus 5.5’s lead holds as other vendors update their models also remains to be seen.

Watching Sol’s Missing Settings and October Releases

Immediate next steps include Artificial Analysis publishing low and max effort settings for GPT-6.1 Sol, which could change its value calculation, and real-world shadow testing of Sol as a routine reviewer outside benchmark conditions. Meyer’s monthly cadence means an October 2026 update is expected, which will test whether the tight score clustering — and the 100x cost spread — persists as vendors respond. Readers following his method should expect the builder/reviewer split to be re-evaluated against new releases each month.

Key Questions

Which AI model does Thorsten Meyer use as his main model in September 2026?

Claude Opus 5.5 at high effort (54 index points, $1.82 per task) for most development work, and at xhigh (56 points, $3.46) for hard problems such as architecture, migrations and trust boundaries. He considers max effort rarely worth the cost.

What is GPT-6.1 Sol used for?

Released 29 September 2026, Sol serves as Meyer’s model for details and review — deep dives into specific files or diffs and independent review passes — at $0.32 to $0.39 per task. He does not use it as a builder, noting Opus 5.5 leads it by 5 index points at xhigh.

Why does Meyer say cost per task matters more than benchmark scores now?

Because six frontier models sit within about 20 index points of each other while cost per task differs by roughly 100 times, the practical question becomes which model clears a quality bar at the lowest cost, rather than which model scores highest.

What are the drawbacks of GPT-6.1 Sol?

Per Meyer, its high and xhigh settings take 57 to 69 seconds to produce a first token, making it unsuitable for interactive use at those levels. Artificial Analysis has not yet published its low or max settings, and one index point of difference is within benchmark noise.

What does ‘effort is not capability’ mean in Meyer’s framework?

It is one of his four working rules: raising the effort setting increases cost and output length but does not make a model smarter, and more effort cannot fill in missing requirements. On Opus 5.5, going from xhigh to max adds 2 index points for 73% more cost per task.

Source: ThorstenMeyerAI.com