Inside My September 2026 AI Stack: Opus Builds, Sol Digs, Jev Decides

🔍 Read the full analysis: Inside My September 2026 AI Stack: Opus Builds, Sol Digs, Jev Decides 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

AI practitioner Thorsten Meyer published his working AI stack on 29 September 2026, the same day OpenAI’s GPT-6.1 Sol launched, arguing that the frontier has shifted from a capability leaderboard to a price curve: six leading models now sit within about 20 index points of each other on the Artificial Analysis Intelligence Index while their cost per task differs by roughly 100x. His answer is a division of labor — Opus 5.5 builds, GPT-6.1 Sol digs and reviews, and Jev decides — built around the question of which model clears a quality bar at the lowest cost per task.

Meyer’s stack assigns Claude Opus 5.5, released 22 September, as the main builder at high or xhigh effort settings. High delivers an index score of 54 for $1.82 per task; xhigh adds 2 points for $3.46 and is reserved for architecture, migrations and trust boundaries. Opus 5.5 holds the highest index score in the field at 58 on its max setting, and — according to Meyer’s data — outscores its more expensive sibling Fable 5.1 by 5 points while costing less per task.

GPT-6.1 Sol, released 29 September at $2/$10 per 1M tokens, is the stack’s detail and review layer. At xhigh it scores 51 for $0.39 per task — roughly one-eighth of GPT-6 Astra’s per-task cost and one-twentieth of Fable 5.1’s, for a score only 1 to 2 points lower. Meyer notes it is unusually concise, with the high setting using 25M output tokens on the index against a median of 82M for comparable models. The tradeoff: high and xhigh settings take 57 to 69 seconds to first token, making it unsuitable for interactive use at those settings.

Effort settings emerge as the biggest cost lever. Meyer’s figures show Opus 5.5 moving from xhigh to max adds 2 index points for 73% more cost per task; medium to max raises cost 4.46x for 7 points. Sonnet 5.5 illustrates the ceiling: at max it writes about 193k output tokens per task — the most Artificial Analysis has measured, per Meyer — and costs $7.60 for 56 points, more than Opus at max. GPT-6 Luna at $0.07 per task handles classification, extraction and routing. Jev, a decision model that cannot write prose, takes over high-volume yes/no and routing judgements.

Why the Price Curve Changes Model Selection

The piece documents a shift many practitioners are making as model capability converges: when scores cluster within 20 points, per-task cost becomes the deciding variable. Sol’s pricing makes an independent review pass affordable on every meaningful change — a different model family checking Opus’s output costs $0.32 to $0.39 per task, which Meyer argues is a better check than a model reviewing itself.

Meyer also pushes back on two common assumptions. First, that effort settings equal capability: “Turning up the dial does not make a model smarter” — more effort cannot fill in missing requirements. Second, that cheaper tokens mean cheaper work: halving model price saves only about 12.5% of real cost in his illustrative example, and a single extra minute of human review erases the saving. He adds four operating rules, including that passing tests are not approval to ship, and that when review fails, the failing case and evidence go back to Opus rather than a generic “try harder.”

The September 2026 Model Field

The stack covers six models released or updated in September 2026: Claude Fable 5.1 (1 September, index 53, $7.63 per task), GPT-6 Astra (3 September, 53, $3.26), Claude Opus 5.5 (22 September, 58, $5.98), GPT-6 Luna (22 September, 37, $0.07), Claude Sonnet 5.5 (28 September, 56, $7.60) and GPT-6.1 Sol (29 September, 51, $0.39). Sol launched at the same token pricing as its week-old GPT-6 Sol predecessor, and even its medium setting matches that earlier model’s score of 48 at one-fifth the per-task cost, per Meyer’s figures.

All scores cited come from the Artificial Analysis Intelligence Index v4.3.x, which Meyer describes as a map of general capability rather than a verdict on any specific workload. Per-1M-token pricing: Opus 5.5 at $4/$20 with cache reads at $0.20; Fable and Astra at $10/$50; Sol at $2/$10; Luna at $0.10/$0.50.

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

— Thorsten Meyer

Caveats in the Index Data

Meyer repeatedly hedges the numbers. He states that one index point is inside the noise, which covers several of the gaps his recommendations turn on. Artificial Analysis has not yet published Sol’s low or max effort settings, so the full cost-quality curve for the newest model is incomplete. The 12.5% saving figure for halved model pricing is explicitly labeled as illustrative, not measured. And the author’s own guidance is that the index is not a verdict on any reader’s workload — he advises shadow-testing before switching models. His role assignments (Astra, Fable, Sonnet 5.5 and Luna as situational alternates rather than defaults) reflect his workload, not a general ranking.

Watching for Sol’s Full Curve

Open questions include Artificial Analysis publishing Sol’s low and max effort settings, which would complete the pricing picture for the newest model. Meyer’s stack itself is dated — he frames it as his answer “today, on 29 September 2026” — and further releases could shift assignments quickly given how tight the score cluster is. For readers, the practical next step Meyer recommends is shadow-testing candidate models against real workloads rather than relying on index scores, particularly before moving high-volume tasks to Sol or Luna on cost grounds.

Key Questions

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

GPT-6.1 Sol is an OpenAI model released on 29 September 2026, priced at $2/$10 per 1M tokens. Per Meyer’s data, its xhigh setting scores 51 on the Artificial Analysis Intelligence Index at $0.39 per task, placing it 1 to 2 points below Astra and Fable 5.1 at a fraction of their cost.

Why does Meyer use Opus 5.5 instead of the highest-effort setting everywhere?

Because effort settings are the dominant cost lever. Going from xhigh to max on Opus 5.5 adds 2 index points for 73% more cost per task, so Meyer reserves xhigh for hard problems like architecture and migrations and runs high (54 points, $1.82) for everyday development.

What are GPT-6.1 Sol’s main drawbacks?

Two: latency and incomplete data. High and xhigh settings take 57 to 69 seconds to first token, ruling out interactive use at those levels, and Artificial Analysis has not yet published low or max settings, so its full curve is unknown.

What is Jev’s role in the stack?

Jev is a decision model that cannot write sentences, used for high-volume yes/no and routing judgements — work where a small, fast model is sufficient and where models like Luna (37 index points, $0.07 per task) also handle classification and extraction.

Should readers copy this stack directly?

Meyer explicitly says no without testing: the Artificial Analysis index is “a map of general capability, not a verdict on your workload,” and he advises shadow-testing before switching anything. His assignments reflect his own development and review work.

Source: ThorstenMeyerAI.com