The AI Tools I Use To Build, Dig, And Decide This September

🔍 Read the full analysis: The AI Tools I Use To Build, Dig, And Decide This September on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 work and review, following Sol’s release on Sept. 29. His comparison of six models argues that reported task costs can differ sharply even when benchmark scores are relatively close, shifting his selection criteria toward the least expensive model that meets a task’s quality bar.

Meyer’s figures are drawn primarily from the Artificial Analysis Intelligence Index v4.3.x, which he describes as a general capability measure rather than a verdict on a particular workload. In his table, Opus 5.5 scores 58 at its top setting and costs $5.98 per task; GPT-6.1 Sol scores 51 at xhigh and costs $0.39. The other entries are Sonnet 5.5 at 56 and $7.60, Fable 5.1 at 53 and $7.63, GPT-6 Astra at 53 and $3.26, and GPT-6 Luna at 37 and $0.07.

Those cost figures are reported per task in the source. Meyer also lists token prices: Opus 5.5 at $4 per million input tokens and $20 per million output tokens, GPT-6.1 Sol at $2 and $10, and Luna at $0.10 and $0.50. He says Opus at high effort scores 54 for $1.82 per task, while xhigh scores 56 for $3.46. At max, it scores 58 for $5.98. These figures illustrate how effort settings change his reported costs as well as the scores.

For Sol, the source lists three settings: medium at 48 and $0.21 per task, high at 50 and $0.32, and xhigh at 51 and $0.39. Meyer says high and xhigh took 57 and 69 seconds, respectively, to produce a first token in the index results. He assigns Sol to reviewing specific files or code changes and examining details, while using Opus for features, APIs, multi-file work, and refactoring. He says Sonnet, Luna, Astra, and Fable serve as alternatives for narrower tasks or second opinions.

Why Task Costs Shape Model Choice

Meyer’s account reflects a practical change in how he evaluates AI tools: a top benchmark score is only one factor if a lower-cost model can meet the needs of a particular task. The reported gap between Opus 5.5 and Sol is relevant to teams that run repeated review passes, where per-task expenses can accumulate. However, the source does not provide a standardized measure of the models’ performance on Meyer’s own development work, so the cost comparisons alone cannot establish which setup saves money overall.

His workflow also gives a separate model a review role. Meyer argues that a model from a different family can offer a useful second perspective on Opus-generated work. He says he uses Sol routinely because its reported cost is low enough for frequent checks. That is his operating choice, not evidence that a second model will catch every error: he also warns that reviewers can share the same flawed requirements, and that passing tests should not be treated as permission to ship.

The September Model Lineup

The source describes a series of releases across September 2026: Fable 5.1 on Sept. 1, GPT-6 Astra on Sept. 3, Opus 5.5 on Sept. 22, Luna on Sept. 22, Sonnet 5.5 on Sept. 28, and GPT-6.1 Sol on Sept. 29. Meyer says six models fall within roughly 20 points on the index while their reported task costs vary by about 100 times. The index comparisons are snapshots, and the source cautions that scores do not substitute for testing against a user’s own workload.

Meyer further reports that Opus 5.5’s xhigh setting adds two index points over high at a higher listed task cost. He says Sonnet 5.5’s max setting costs $7.60 per task for a score of 56, while its high setting costs $1.08 for a score of 47. His preferred settings therefore vary by job: high or xhigh for development with Opus, and medium for documents and routine work. These are recommendations based on his reading of the index and his stated workflow.

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

— Thorsten Meyer

What the Benchmarks Cannot Show

The source does not provide the underlying task definitions, measurement methods, or confidence intervals for the reported per-task costs, so readers cannot assess how closely those figures match their own use. Meyer notes that a one-point index difference is within the noise and says low and max results for GPT-6.1 Sol had not been published at the time of writing. The account also does not establish whether Sol’s reported first-token delays would apply across different prompts or service conditions.

Most importantly, the index is a general benchmark. The source gives no independent comparison of how often each model produces correct code, catches defects, or reduces human review time on a shared set of real projects. The article’s workflow and cost judgments are Meyer’s reported choices; performance on other workloads remains to be tested.

Test Models Against Your Work

Meyer recommends shadow-testing a candidate model before switching a workflow. That means checking it against the work it would actually handle and comparing the results with the existing process. For his own setup, the next step is continued use of Opus for building and Sol for review, with Astra or Fable as alternatives when those models prove more useful in a specific comparison. No later benchmark release or change to his stack is specified in the source.

Key Questions

Which models does Thorsten Meyer use for building and review?

Meyer says Opus 5.5 is his main model for building, while GPT-6.1 Sol handles detailed investigation and review. He uses other models for specific tasks or as alternatives.

Why does Meyer use GPT-6.1 Sol for reviews?

He reports that Sol costs $0.32 per task at high and $0.39 at xhigh in the cited index results. He says that makes a separate review pass affordable to run routinely. The source does not show that this cost or performance will apply to every user’s workload.

Does the benchmark identify the best model for every task?

No. Meyer says the Artificial Analysis Intelligence Index measures general capability and advises readers to test models on their own work before changing systems.

What is still missing from the GPT-6.1 Sol results?

At the time of Meyer’s Sept. 29 account, he said the index had not published Sol’s low or max settings. He also notes that the reported high and xhigh settings had long first-token times in the index, while real-world timing may depend on the task and service conditions.

Source: ThorstenMeyerAI.com