September 2026 AI: The Stack I Use To Build, Dig, And Decide

🔍 Read the full analysis: September 2026 AI: The Stack I Use To Build, Dig, And Decide on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

GPT-6.1 Sol, released on 29 September 2026, completes a four-week stretch in which six frontier AI models landed within about 20 index points of each other on capability benchmarks while their cost per task differs by roughly 100x — a shift that turns model selection from a leaderboard question into a price question, according to practitioner analysis published by Thorsten Meyer. Meyer’s recommended stack pairs Claude Opus 5.5 as the main builder with the newly launched Sol as a low-cost reviewer at $0.39 per task.

All capability figures 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. At the top, Claude Opus 5.5 (released 22 September) scores 58 at its max effort setting, costing $5.98 per task — about 17 tasks per $100. At the bottom, GPT-6 Luna (22 September) scores 37 at $0.07 per task, or roughly 1,429 tasks per $100, positioning it for classification, extraction and routing work.

Three findings stand out in Meyer’s comparison. First, Opus 5.5 outscores its more expensive sibling Claude Fable 5.1 by 5 points while costing less per task. Second, Claude Sonnet 5.5 at max effort costs more per task than Opus at max while scoring 2 points lower. Third, GPT-6.1 Sol costs about one-eighth of GPT-6 Astra and one-twentieth of Fable per task for a score only 1 to 2 points lower.

Meyer reports that the effort setting, not model choice, is the dominant cost lever. On Opus 5.5, moving from medium to max effort raises cost 4.46x for 7 additional index points; xhigh-to-max adds 2 points for 73% more cost. He runs Opus at high (54 points, $1.82 per task) for everyday development and xhigh (56 points, $3.46) for architecture, migrations and trust boundaries.

Why Routine AI Review Became Affordable

The core practical change is that independent second-opinion review is now cheap enough to run on every meaningful change. Meyer’s argument: a model from a different family reviewing Opus output is a stronger check than Opus reviewing itself, and at $0.32 to $0.39 per task, review passes can be routine rather than exceptional.

Meyer also argues the economics reframe cost-cutting: halving model price saves only 12.5% of real cost in his illustrative example, and a single extra minute of human review erases the saving — so cheap tokens alone do not mean cheap work. He maintains four operating rules, including that effort settings do not make a model smarter, that a second model reading the same flawed spec is not an independent review, and that failing cases should be handed to the builder model with evidence, not just instructions to “try harder”.

September’s Six-Model Release Wave

The month’s releases, per Meyer’s table: Claude Fable 5.1 (1 September, index 53, $7.63 per task), GPT-6 Astra (3 September, 53, $3.26), Opus 5.5 (22 September, 58, $5.98), GPT-6 Luna (22 September, 37, $0.07), Sonnet 5.5 (28 September, 56, $7.60) and GPT-6.1 Sol (29 September, 51 at xhigh, $0.39). Per-million-token prices: Opus 5.5 at $4 input / $20 output, Fable and Astra at $10 / $50, Sol at $2 / $10, and Luna at $0.10 / $0.50.

Sol launched at the same list price as its week-old predecessor, and per Artificial Analysis figures, even its medium setting matches the earlier GPT-6 Sol’s score of 48 at one-fifth the per-task cost. Meyer notes Sol is unusually concise — its high setting used about 25M output tokens on the index against a median of 82M for comparable models.

“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 100x.”

— Thorsten Meyer, ThorstenMeyerAI.com

Benchmark Noise and Sol’s Missing Settings

Several limitations remain. Artificial Analysis has not yet published low or max effort settings for GPT-6.1 Sol, and Meyer notes that a single index point falls inside measurement noise — relevant given Sol trails Astra and Fable by only 1 to 2 points. Sol’s high and xhigh settings take 57 to 69 seconds to produce a first token, which Meyer says rules it out as an interactive model at those settings.

Meyer’s cost-of-work example — that halving model price saves 12.5% of real cost — is explicitly labeled illustrative, not measured. His stack recommendations are a single practitioner’s workflow, not a controlled evaluation, and the piece itself advises shadow-testing before switching models.

Watching for Sol’s Full Effort Curve

The near-term developments to watch are Artificial Analysis publishing Sol’s low and max effort settings, which would complete the price-performance picture for the new model, and any pricing moves from Anthropic or Google in response to Sol’s aggressive per-task cost. Meyer’s stated workflow will continue evolving as new models land — his own framing treats the current stack as a snapshot dated 29 September 2026, not a stable recommendation.

For teams adopting a similar split, the practical next step Meyer implies is shadow-testing: running candidate models on real workloads before committing, rather than relying on index scores alone.

Key Questions

What is GPT-6.1 Sol, and when was it released?

GPT-6.1 Sol is a frontier model released on 29 September 2026 at $2 input / $10 output per million tokens — the same list price as its week-old predecessor. It scores 51 on the Artificial Analysis Intelligence Index v4.3.x at xhigh effort, at $0.39 per task.

Why pair Opus 5.5 with GPT-6.1 Sol instead of using one model?

According to Meyer, Opus 5.5 leads on capability (56 points at xhigh) for building work, while Sol’s low per-task cost ($0.32–$0.39) makes routine independent review affordable. A reviewer from a different model family is also a stronger check than a model reviewing its own output.

Are these benchmark scores definitive?

No. All figures come from the Artificial Analysis Intelligence Index v4.3.x, which Meyer describes as a map of general capability rather than a verdict on individual workloads. He recommends shadow-testing any model on your own tasks before switching, and notes one index point is within noise.

What are GPT-6.1 Sol’s main drawbacks?

Per Meyer’s data, Sol’s high and xhigh settings take 57 to 69 seconds to first token, making it unsuitable for interactive use at those settings. It also trails Opus 5.5 by 5 index points at xhigh, and its low and max effort settings have not yet been benchmarked.

Does a cheaper model mean cheaper overall work?

Not necessarily. In an illustrative example, Meyer calculates that halving model price saves only 12.5% of real cost, which a single extra minute of human review can erase. He argues task outcomes, not token prices, determine true cost.

Source: ThorstenMeyerAI.com