Building, Researching, Deciding: My AI Setup For September

🔍 Read the full analysis: Building, Researching, Deciding: My AI Setup For September on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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

Thorsten Meyer says he now uses Claude Opus 5.5 as his main model for building and GPT-6.1 Sol for detailed research and review, in a September 29 account that puts cost per task at the center of his AI choices. The change reflects his comparison of six models whose reported Intelligence Index scores are relatively close while their task costs vary widely.

Meyer bases the comparison on the Artificial Analysis Intelligence Index v4.3.x, describing it as a measure of general capability rather than a verdict on any particular workload. In his table, Opus 5.5 scores 58 at its top setting and costs $5.98 per task; GPT-6.1 Sol at xhigh scores 51 and costs $0.39. The other listed models are Sonnet 5.5, Fable 5.1, GPT-6 Astra and GPT-6 Luna. Meyer says the figures should be shadow-tested against a user’s own work before changing a workflow.

His assignments reflect both those scores and prices. He uses Opus 5.5 at high effort for regular development, and xhigh for demanding work such as architecture, migrations and trust boundaries. He assigns Sol at high or xhigh to inspect specific files and diffs and to provide a second review. Astra or Fable are alternatives when Sol and Opus disagree; Sonnet handles scoped subtasks, while Luna is reserved for routine checks and bulk classification.

The source reports that GPT-6.1 Sol launched on September 29 at the same listed token prices as its predecessor, GPT-6 Sol: $2 per million input tokens and $10 per million output tokens. On the index, Sol scores 48 at medium, 50 at high and 51 at xhigh, with reported costs of $0.21, $0.32 and $0.39 per task, respectively. At high and xhigh, its reported time to first token is 57 and 69 seconds. The source says Artificial Analysis has not yet published low or max results for Sol.

Why Review Costs Change the Workflow

Meyer’s account highlights a practical trade-off for people who use AI across many work tasks: a small score difference may come with a large cost difference. He argues that Sol’s reported cost makes it affordable to use as a routine second reviewer, while a more expensive model can remain assigned to work where its higher score is useful. That is his workflow judgment, rather than proof that the same allocation suits every team.

He also emphasizes that the effort setting can materially affect the bill. In the figures he cites, Opus 5.5 rises from $1.82 per task at high to $3.46 at xhigh and $5.98 at max. The top setting adds two index points over xhigh, according to the table. For Sonnet 5.5, the listed max setting costs $7.60 per task for a score of 56, compared with $2.74 and a score of 52 at xhigh. Those comparisons make model selection only part of the spending decision.

The approach also gives a model from a different family a review role. Meyer says that a separate model reviewing Opus can provide a useful check, but cautions that different models reading the same incomplete requirements may share the same blind spot. He says passing tests alone should not be treated as approval to ship.

How the September Model Comparison Works

The account is a snapshot dated September 29, 2026, not a permanent ranking. Meyer says the AI field has shifted, in his view, from a leaderboard question—“which model is smartest?”—to a cost and capability question: which model meets a task’s quality bar at the lowest cost. His comparison covers six named models and uses each model’s reported top setting for its headline table, with additional effort-level figures for Opus and Sonnet.

In that table, Opus 5.5 is listed at 58 points, ahead of Fable 5.1 at 53, while costing less per task in the cited figures. Sonnet 5.5 is listed at 56 at max effort, but its cost rises to $7.60. Sol’s xhigh score is 51, compared with Astra at 53 at max and Fable at 53. Meyer notes that a one-point difference may fall within measurement noise. He also reports that Sol’s output-token count at high was 25 million in the index, against a median of 82 million for comparable models; that index-specific figure does not establish how much output a user’s own tasks will require.

The source includes token prices as well as task costs. It lists Opus at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20; Fable and Astra at $10 and $50; Sol at $2 and $10; and Luna at $0.10 and $0.50. Token rates and cost per task are different measures: the latter depends on the work and the evaluation method. The source’s final cost discussion is incomplete, so its illustrative calculation about human review time cannot be fully assessed from the supplied material.

““Which model clears my quality bar at the lowest cost per task?””

— Thorsten Meyer

What the Index Cannot Establish

The figures are attributed to Artificial Analysis, but the supplied source does not include a link to the underlying index entry, its full methodology or the tasks used to calculate cost per task. It is therefore unclear how directly the reported prices and scores map to a reader’s own prompts, usage patterns or quality requirements. Meyer himself says to shadow-test before switching.

Some comparisons remain provisional: the source says Sol’s low and max effort results had not been published at the time, and says a one-point score difference is within the noise. The text also ends during an illustrative explanation of how human review time can affect savings. It does not provide the rest of that calculation or enough detail to treat it as a measured result. The reported model prices, availability and index scores may also change after this dated snapshot.

Testing Before Changing Defaults

Meyer’s stated next step for anyone considering a similar setup is to compare candidate models on their own work before switching defaults. His proposed workflow uses a second model to review meaningful changes and sends failed cases, alongside evidence, back to the builder for follow-up. The source does not identify a scheduled update, a future benchmark release or a date when Meyer plans to revise these assignments.

Readers applying the comparison can track whether each model meets their quality bar, how long it takes to respond and what a completed task costs at the chosen effort level. Those local results would show whether the September 29 allocation remains useful for their workloads as models and published evaluations change.

Key Questions

Which models does Meyer use for building and review?

He says he uses Opus 5.5 as his main builder and GPT-6.1 Sol at high or xhigh for detailed research and review.

What does the source report GPT-6.1 Sol costs?

The cited index lists Sol at $0.32 per task at high and $0.39 at xhigh. These are reported task costs in the index, not a guaranteed price for every user’s work.

Does the index identify the best model for every workload?

No. Meyer describes it as a map of general capability and recommends testing models against the work they would actually handle.

Why does Meyer use more than one model?

He assigns models to different jobs based on their reported capability and cost. He says a second model from another family can review Opus’s output, while warning that separate models can still share a blind spot if they receive the same flawed requirements.

Source: ThorstenMeyerAI.com