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

🔍 Read the full analysis: Opus Builds, Sol Digs, Jev Decides: My September 2026 AI Stack on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer, an independent AI commentator, published his working model stack on 29 September 2026, arguing that the frontier AI market has shifted from a capability race to a price curve: six leading 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 100x. His conclusion for practitioners: the question is no longer “which model is smartest?” but “which model clears my quality bar at the lowest cost per task?” His answer pairs Claude Opus 5.5 for building with the just-released GPT-6.1 Sol for review, plus a decision-only model called Jev for high-volume yes/no routing.

Meyer’s stack assigns specific roles by cost and capability. Opus 5.5, released 22 September, is his main builder: at the “high” effort setting it scores 54 index points for $1.82 per task, and at “xhigh” it reaches 56 points for $3.46 — the setting he reserves for architecture, migrations and trust boundaries. GPT-6.1 Sol, released the day of publication, handles details and review at $0.32 to $0.39 per task at its high and xhigh settings. Sonnet 5.5 (high) and GPT-6 Luna cover scoped subtasks and bulk classification, while Astra and Fable serve only as tie-breaking second opinions.

Three findings anchor the piece, all drawn from the Artificial Analysis index. First, Opus 5.5 outscores its more expensive sibling Fable 5.1 by 5 points while costing less per task. Second, Sonnet 5.5 at maximum effort costs more per task than Opus at maximum while scoring 2 points lower, which Meyer says makes that setting hard to justify. Third, 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.

The effort setting, not model choice, is the biggest cost lever, according to Meyer’s numbers. On Opus 5.5, moving from xhigh to max adds 2 index points for 73% more cost per task; from medium to max, cost rises 4.46x for 7 points. Sol has real limitations: at high and xhigh it takes 57 to 69 seconds to produce a first token, making it unsuitable for interactive use, and Opus 5.5 still leads it by 5 points at xhigh.

Why Cross-Family Review at $0.39 Matters

The piece’s central argument is economic rather than technical: once capability gaps between frontier models shrink to single index points, cost per task becomes the primary decision variable. At $0.39 per task, a review pass is cheap enough to run on every meaningful code change — something that was not practical when comparable models cost $3 to $8 per task.

Meyer also argues that a different model family reviewing Opus output is a stronger check than Opus reviewing itself. This cross-family review pattern, he suggests, is the highest-value use of the new cheap-but-capable tier. He cautions against complacency with four rules: effort is not capability; a different model reading the same flawed spec is not an independent review; passing tests are not approval to ship; and failing cases should be handed back with evidence, not just an instruction to “try harder.”

His cost framing also pushes back on token-price obsession: halving model price saves only 12.5% of real task cost in his illustrative example, and a single extra minute of human review can erase the saving.

A Month of Crammed Frontier Releases

September 2026 was unusually dense for frontier releases, according to the timeline in Meyer’s piece: Fable 5.1 on 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. Sol launched at the same list price as its week-old predecessor ($2 input / $10 output per 1M tokens), and Artificial Analysis already lists three effort settings for it, with its medium setting matching GPT-6 Sol’s score of 48 at one-fifth the per-task cost.

Meyer notes that Sonnet 5.5 at max effort writes about 193k output tokens per task, the most Artificial Analysis has measured, which explains why its cost jumps from $2.74 to $7.60 for only 4 additional index points. Sol shows the opposite profile: its high setting used 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.”

— Thorsten Meyer

What the Index Doesn’t Tell You

Meyer is explicit that all capability scores come from the Artificial Analysis Intelligence Index v4.3.x, which he describes as a map of general capability rather than a verdict on any specific workload. He advises shadow-testing before switching models in production.

Several data points remain incomplete. Artificial Analysis has not yet published low or max effort settings for GPT-6.1 Sol, and Meyer notes that a single index point is within measurement noise. His cost-savings example — halving model price saving 12.5% of real cost — is described as illustrative, not measured. The piece also gives no detail on the Jev decision model’s benchmarks or pricing, and the “Jev decides” role is asserted without supporting numbers. Real-world performance on individual workloads may differ substantially from index results.

Benchmarks and Settings Still Landing

Expect Artificial Analysis to publish the missing low and max effort settings for GPT-6.1 Sol, which could change its value calculation. The compressed release cadence — three major models in the final week of September alone — suggests further frontier releases in October, and Meyer’s stated method is to re-run the quality-bar-versus-cost analysis each time the price curve shifts. For teams adopting a similar stack, his stated next step is unglamorous: shadow-test candidate models against real workloads before switching defaults, and verify that cross-family review actually catches errors that self-review misses.

Key Questions

What is the main claim of Thorsten Meyer’s September 2026 stack?

That frontier models have converged in capability — six models within roughly 20 index points — while task costs differ by about 100x, so model selection should be driven by cost per task at a given quality bar, not by raw benchmark scores.

Why does Meyer use GPT-6.1 Sol for review instead of Opus?

Two reasons, per Meyer: Sol costs $0.32 to $0.39 per task, making routine review affordable, and a different model family provides a more independent check than a model reviewing its own output.

What are GPT-6.1 Sol’s main drawbacks?

At high and xhigh settings it takes 57 to 69 seconds to first token, ruling out interactive use, and it trails Opus 5.5 by 5 index points at xhigh. Low and max effort settings have not yet been published.

What is the effort-setting trade-off on Opus 5.5?

According to Meyer’s figures, going from xhigh to max adds 2 index points for 73% more cost per task, and from medium to max raises cost 4.46x for 7 points. He therefore runs Opus at high or xhigh and treats max as rarely worth it.

Are these recommendations backed by peer-reviewed or systematic evaluation?

No. The scores cited are from the Artificial Analysis Intelligence Index v4.3.x, a third-party benchmark, and the stack reflects one practitioner’s reported workflow. Meyer himself advises shadow-testing before switching anything.

Source: ThorstenMeyerAI.com