🔍 Read the full analysis: A Look At How I Work With AI In September 2026 on ThorstenMeyerAI.com
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
TL;DR
Analyst Thorsten Meyer published his September 2026 AI working setup, built around Claude Opus 5.5 as the main builder and the newly released GPT-6.1 Sol as a low-cost reviewer. With six frontier models within roughly 20 index points but 100x apart in cost per task, he argues price per task now matters more than leaderboard scores.
Independent AI analyst Thorsten Meyer published his working AI setup on 29 September 2026, assigning Claude Opus 5.5 as his main model for building software and the just-launched GPT-6.1 Sol as a cheap second reviewer. His core argument: with six frontier models now clustered within about 20 points on capability indexes while their cost per task differs by roughly 100 times, the practical question has shifted from “which model is smartest?” to “which model clears a quality bar at the lowest cost per task?”
All capability figures in Meyer’s analysis come from the Artificial Analysis Intelligence Index v4.3.x. At the top of the field, Claude Opus 5.5 (released 22 September) scores 58 at its max setting at $5.98 per task, or 17 tasks per $100. At the bottom, GPT-6 Luna (also 22 September) scores 37 at $0.07 per task — 1,429 tasks per $100 — which Meyer assigns to classification, extraction and routing. In between sit Claude Sonnet 5.5 (56, $7.60), Claude 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. 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 max effort costs more than Opus at max for 2 fewer points, which Meyer says disqualifies it at that setting. Third, GPT-6.1 Sol costs about one-eighth of Astra and one-twentieth of Fable per task for a score only 1 to 2 points lower.
Meyer also identifies the effort setting — not model choice — as the biggest 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. He therefore runs Opus at high (54 points, $1.82 per task) for everyday development and xhigh (56 points, $3.46) only for hard problems such as architecture, migrations and trust boundaries.
Opus builds. Sol reviews. Jev decides.
One price tape, six models
Score against cost, at every effort setting
The effort dial moves the bill more than the model
Claude Opus 5.5
Claude Sonnet 5.5
GPT-6.1 Sol: near-Astra scores at a fraction of the price
Three published settings
| Setting | Index | Cost per task | Output tokens | First token |
|---|---|---|---|---|
| medium | 48 | $0.21 | 15M | 5.3 s |
| high | 50 | $0.32 | 25M | 57 s |
| xhigh | 51 | $0.39 | 36M | 69 s |
Same score band, very different bill
My stack: who builds, who reviews
Cheaper tokens are not cheaper work
Read the numbers with four warnings
Part 2: Jev, the model that decides instead of writing
One call in, typed answers out
Three question types
Confidence is the superpower
Three uses running in my publishing operation
The fit test, then the shadow test
- Replay 300 to 500 past decisions
- Compare overall and per confidence band
- Read 20 disagreements, decide who was right
- High band at 95% or better?
- Own flag, off by default
- Canary on 5 to 10 units
- Roll out in the confident band only
24 use cases, sorted by how well they fit
Proven in production
- 1Relevance gate
- 2Language check
- 3Classifier fallback
Publishing and content
- 4Thin-source detector
- 5Same-event dedupe
- 6Product fits roundup
- 7Disclosure present
- 8Headline quality
- 9Comment moderation
Commerce and support
- 10Support-ticket routing
- 11Return-reason coding
- 12Review to feature complaints
- 13Catalogue taxonomy
- 14Order-fraud pre-triage
Software and AI systems
- 15LLM guardrail
- 16RAG passage filter
- 17Citation check
- 18Tool and intent routing
- 19Log-line triage
- 20PR risk triage
Business ops and home
- 21Inbox triage
- 22Expense categorisation
- 23Lead qualification
- 24Smart-home intent
Limits, cost and one hard rule
Why Cost Per Task Now Drives Model Choice
Meyer’s framework matters because it reflects a broader shift in how professional users select AI models. When capability gaps between frontier models shrink to single index points, the marginal value of switching models is small compared with the marginal cost. A model scoring 2 points lower at one-eighth the price becomes viable for routine, high-volume work like code review, where the review pass at $0.32 to $0.39 per task is, in his words, cheap enough to be routine.
His second structural point concerns cross-model review. He argues a different model family reviewing Opus’s output is a better check than Opus reviewing itself, and that affordability makes running an independent reviewer on every meaningful change economically feasible. He pairs this with four working rules: effort is not capability; more effort cannot fill in missing requirements; a different model reading the same flawed spec is not an independent review; and passing tests are not approval to ship.
He also cautions against over-optimizing price, noting that halving model cost saves only 12.5% of real task cost, and that a single extra minute of human review can erase the saving — a figure he presents as illustrative rather than measured.
Products Worth Considering
Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)
LG UltraGear 32GX870A-B 32-inch 4K UHD OLED Gaming Monitor
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
GPT-6.1 Sol’s Launch-Day Numbers
GPT-6.1 Sol launched on 29 September, the same day as Meyer’s write-up, priced identically to its week-old predecessor at $2 per million input tokens and $10 per million output tokens. Artificial Analysis lists three effort levels: medium (48 index, $0.21 per task), high (50, $0.32) and xhigh (51, $0.39). Even its medium setting matches the earlier GPT-6 Sol’s score of 48 at one-fifth of that model’s $1.06 per task.
Sol is unusually concise, using 25 million output tokens on the index at high against a median of 82 million for comparable models. The trade-offs are real: 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 it by 5 points at xhigh.
The wider context is a crowded September release cycle: Fable 5.1 (1 September), Astra (3 September), Opus 5.5 and Luna (22 September), Sonnet 5.5 (28 September) and Sol (29 September) all shipped within four weeks.
“In four weeks, the AI frontier stopped being a leaderboard and became a price curve.”
— Thorsten Meyer
Benchmark Limits and Missing Settings
Several things remain unresolved. Artificial Analysis has not yet published low or max settings for GPT-6.1 Sol, so its full cost curve is unknown. Meyer notes that one index point is inside measurement noise, meaning the exact ordering of closely scored models — Sol, Astra and Fable differ by 1 to 2 points — should not be treated as settled. All scores come from a general-purpose index rather than task-specific benchmarks, and Meyer explicitly recommends shadow-testing before switching models. His cost-saving arithmetic about human review time is labeled illustrative, not measured. Whether the current price compression holds as vendors adjust pricing after launch is also unknown.
Watching the Price Curve Settle
Meyer expects to keep Opus 5.5 at high or xhigh as the default builder and Sol as the standing reviewer, adjusting only when a workload’s own tests justify it. Near-term developments worth watching include Artificial Analysis publishing low and max settings for GPT-6.1 Sol, which would complete its cost picture; any pricing response from Anthropic or OpenAI to Sol’s aggressive per-task economics; and whether the 100x cost spread across the frontier narrows as new models ship in October.
Key Questions
What is Thorsten Meyer’s main AI model setup as of September 2026?
He uses Claude Opus 5.5 at high or xhigh effort as his main model for building, GPT-6.1 Sol at high or xhigh for details and independent review, and Sonnet 5.5, Astra, Fable and Luna as targeted alternates for scoped subtasks, second opinions and bulk classification.
Why does he prefer GPT-6.1 Sol for review instead of Opus?
Sol scores only 1 to 2 points below Astra and Fable but costs roughly one-eighth to one-twentieth as much per task ($0.32 to $0.39), making a review pass on every meaningful change affordable. He also argues a different model family provides a better check than a model reviewing itself.
What are the drawbacks of GPT-6.1 Sol?
At high and xhigh settings it takes 57 to 69 seconds to produce a first token, so it is not suited to interactive use. It also trails Opus 5.5 by 5 index points at xhigh, and its low and max settings have not yet been benchmarked.
Why does effort level matter more than model choice?
On Opus 5.5, going from medium to max effort raises cost 4.46 times for 7 index points, a larger cost swing than switching between most comparable models. Meyer reserves max settings for rare cases and treats medium as the default for routine document work.
How reliable are the index scores cited?
All scores 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. He notes that one index point falls inside measurement noise and recommends shadow-testing before switching models.
Source: ThorstenMeyerAI.com
