How I Use Three AI Tools In September 2026

🔍 Read the full analysis: How I Use Three AI Tools In September 2026 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

On September 29, 2026, Thorsten Meyer described using Claude Opus 5.5 to build, GPT-6.1 Sol to examine details and review work, and Jev for high-volume yes-or-no and routing decisions. His account argues that reported task costs can differ sharply among models with relatively close benchmark scores, a distinction that could affect how teams assign routine and demanding work.

Meyer says Opus 5.5 at high or xhigh effort is his main model for features, APIs, multi-file work and refactoring. He reserves xhigh for difficult tasks such as architecture, migrations and trust boundaries. In the Artificial Analysis Intelligence Index v4.3.x figures he cites, Opus scores 54 at high for $1.82 per task and 56 at xhigh for $3.46. Its max setting scores 58 but costs $5.98 per task.

He assigns GPT-6.1 Sol detailed file and code-diff work, along with review of Opus’s output. The index figures he quotes list Sol at medium for $0.21 per task, high for $0.32 and xhigh for $0.39. At high and xhigh, the reported time to first token is 57 and 69 seconds, respectively, which Meyer says makes those settings less suited to interactive use. He calls a separate model family a useful second pair of eyes, while warning that different models can still share a flawed specification.

For other work, Meyer names Jev as a decision model for routing and binary judgments, and lists Astra, Fable, Sonnet and Luna as alternatives for selected jobs. He says Sonnet’s high setting suits scoped work and documents, while Luna suits classification and extraction. His proposed practice is to test models against the intended workload before switching; the source does not provide those shadow-test results.

How Meyer Splits Model Work

The account offers a concrete example of assigning work by task cost and role, rather than choosing a single model for every request. In Meyer’s workflow, a more expensive model handles complex implementation, while a lower-cost model can review meaningful changes routinely. If the cited costs hold for a team’s own tasks, that division could make an additional review pass more practical.

Those implications depend on the task. The index is a measure of general capability, Meyer says, and does not establish which model performs best on a particular team’s code, documents or routing decisions. The article presents his operating choices and benchmark interpretation; it does not report a controlled comparison of his full workflow or measured savings across a team.

The Index Behind His Choices

Meyer bases the model comparisons on the Artificial Analysis Intelligence Index v4.3.x, while noting that it maps general capability rather than settling performance for an individual workload. The cited top-setting scores range from 37 for GPT-6 Luna to 58 for Claude Opus 5.5. The source reports costs per task ranging from $0.07 for Luna to $7.63 for Claude Fable 5.1, but does not explain in the supplied material how each task’s token use was standardized.

His comparison also focuses on effort settings, which change both scores and costs. For Opus 5.5, the source lists $0.55 per task at low effort, $1.34 at medium, $1.82 at high, $3.46 at xhigh and $5.98 at max. Meyer favors high or xhigh for building and says max is rarely worthwhile for his work. GPT-6.1 Sol launched on September 29, according to the source, at the same $2 per million input tokens and $10 per million output tokens as its predecessor.

“The practical reading: Sol is not the model I ask to build. It is the model I can afford to run on everything.”

— Thorsten Meyer

What the Benchmarks Cannot Show

The supplied source does not establish how well these assignments perform across other organizations or workloads. It does not include results from Meyer’s recommended shadow tests, a measured comparison of review accuracy, or an observed total-cost analysis that includes human time. The cost-per-task figures are attributed to the index, but the material provided does not detail the task mix behind them.

Some index settings were also unavailable at publication: Meyer says low and max results for GPT-6.1 Sol had not yet been published. He cautions that a one-point score difference may fall within measurement noise. The source also gives no operational details or benchmark results for Jev, and its final cost example is cut off, so it cannot support a conclusion about total savings after human review.

Test Against Real Workloads

Meyer recommends shadow-testing models on a team’s own tasks before changing assignments. That would let teams compare output quality, review needs, response time and cost under their specific conditions. The source gives no date for further index results or a planned follow-up, so whether new GPT-6.1 Sol effort settings or additional measurements will change Meyer’s choices remains open.

Key Questions

Which three tools does Meyer use most?

He identifies Claude Opus 5.5 for building, GPT-6.1 Sol for detailed work and review, and Jev for high-volume yes-or-no and routing judgments.

Why does Meyer use GPT-6.1 Sol for review?

He says its cited cost of $0.32 to $0.39 per task at high or xhigh makes routine review affordable in his workflow. He also values a reviewer from a different model family, while acknowledging that this does not remove flaws in the shared specification.

Does the index prove which model is best for a team?

No. Meyer describes the Artificial Analysis Intelligence Index as a map of general capability, not a verdict on an individual workload. He recommends testing models against the team’s own tasks.

What remains unknown about Meyer’s approach?

The source does not provide results from his shadow tests, a measured comparison of review quality, or complete cost accounting that includes human review time. It also says some GPT-6.1 Sol effort settings had not yet been published.

Source: ThorstenMeyerAI.com