How I Match AI Tools To Tasks In My September Stack

🔍 Read the full analysis: How I Match AI Tools To Tasks In My September Stack on ThorstenMeyerAI.com

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

Get smart everyday buys 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

Practitioner Thorsten Meyer published his September 2026 AI tool stack on 29 September, built around a simple shift: with six frontier models now clustered within roughly 20 index points of each other on general capability while their cost per task differs by about 100 times, model choice has moved from “which model is smartest?” to “which model clears my quality bar at the lowest cost per task?” His answer pairs Claude Opus 5.5 as the main builder with the same-day-released GPT-6.1 Sol as a routine second-opinion reviewer at roughly one-tenth the cost per task.

Meyer’s stack assigns each model a role rather than a rank. Opus 5.5 at high or xhigh effort is the main model for building features, APIs and multi-file work. GPT-6.1 Sol, released 29 September, handles detail work and review. Claude Sonnet 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6 Luna serve as alternates for scoped jobs, and a decision model called Jev — which Meyer notes cannot write a sentence — handles high-volume yes/no and routing judgements. All capability 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.

The pricing table at the heart of the piece shows the spread. Opus 5.5 tops the index at 58 points at a cost of $5.98 per task (17 tasks per $100), while GPT-6 Luna sits at 37 points for $0.07 per task (1,429 tasks per $100). GPT-6.1 Sol at xhigh lands at 51 points for $0.39 per task — about one-eighth of Astra’s per-task cost and one-twentieth of Fable’s for a score only 1 to 2 points lower. Meyer highlights three findings: 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; and Sol delivers near-Astra scores at a fraction of the price.

The analysis also identifies effort settings as the dominant cost lever. On Opus 5.5, moving from xhigh to max adds 2 index points but 73% more cost per task; going from medium to max raises cost 4.46 times for 7 points. Meyer 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, calling max “rarely worth it.” Sonnet 5.5 at max writes about 193k output tokens per task — the most Artificial Analysis has measured, according to Meyer — driving its cost from $2.74 to $7.60 for 4 additional points.

Why Routine AI Review Became Affordable

The core argument is that cheap review passes change workflow economics. Meyer writes that a review pass at $0.32 to $0.39 per task is “cheap enough to be routine,” letting him run a second model over every meaningful change. He argues a different model family reviewing Opus’s output is a better check than Opus reviewing itself — though he adds the caveat that a different model is not an independent review if both read the same flawed specification.

The piece also pushes back on price-focused model selection. Meyer’s illustrative example: halving model price saves only about 12.5% of real cost, and a single extra minute of human review can erase that saving — a point he flags as illustrative rather than measured. Four operating rules anchor the stack: effort is not capability; more effort cannot fill in missing requirements; passing tests are not approval to ship; and failed reviews should hand the failing case and evidence to the builder model, not just “try harder.”

GPT-6.1 Sol’s Same-Day Launch

GPT-6.1 Sol launched on 29 September at the same published prices as its roughly week-old predecessor — $2 per 1M input tokens and $10 per 1M output tokens — and Artificial Analysis already lists three effort settings for it. Even at medium, Sol matches the earlier GPT-6 Sol’s score of 48 at one-fifth of that model’s $1.06 per-task cost. At xhigh it reaches 51 points with 36M output tokens on the index, against a median of 82M for comparable models, which Meyer reads as unusual conciseness.

The catches he notes are latency and a capability gap: high and xhigh settings take 57 to 69 seconds to produce a first token, ruling out interactive use at those settings, and Opus 5.5 still leads Sol by 5 points at xhigh (56 versus 51). Artificial Analysis has not yet published low or max settings for Sol, and Meyer cautions that one index point is inside measurement noise.

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

— Thorsten Meyer

Benchmark Limits and Unmeasured Settings

Several limits remain. The Artificial Analysis index measures general capability, not any specific workload, and Meyer explicitly advises shadow-testing before switching models. Sol’s low and max effort settings have not been published, so its full cost curve is unknown. The one-eighth and one-twentieth cost comparisons rest on the index’s task-cost methodology, which may not reflect real production tasks. Meyer’s 12.5% saving figure is labeled illustrative, not measured. And the author’s role assignments — Jev for routing, Luna for classification — are his own workflow choices, not benchmark-validated recommendations.

Watching for Sol’s Missing Benchmarks

The immediate open item is Artificial Analysis publishing GPT-6.1 Sol’s low and max effort settings, which would complete its cost-quality curve. Meyer’s stated practice is to re-run this matching exercise as new releases land — the piece itself is the September update of a recurring review — and readers following the framework should expect an October revision if October releases shift the price-performance frontier. He also frames cross-model review as an evolving practice, with the standing question of how independent two models truly are when they share the same input specification.

Key Questions

Which model does Meyer use as his primary builder?

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

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

Two reasons, per Meyer: Sol’s review passes cost $0.32 to $0.39 per task, cheap enough to run routinely, and a reviewer from a different model family provides a better check than a model reviewing its own output.

What are GPT-6.1 Sol’s main drawbacks?

Its high and xhigh settings take 57 to 69 seconds to produce a first token, making it unsuitable for interactive use, and Opus 5.5 still outscores it by 5 points at xhigh. Its low and max settings had not been benchmarked at publication time.

Does a higher effort setting make a model smarter?

No, according to Meyer. Effort raises cost and can add a few index points, but he states that “effort is not capability” and that more effort cannot compensate for missing requirements.

Should readers copy this stack directly?

Meyer advises against it. The Artificial Analysis index measures general capability, not individual workloads, and he recommends shadow-testing any candidate model on your own tasks before switching.

Source: ThorstenMeyerAI.com