Inside My AI Workflow: Building, Digging, And Deciding

🔍 Read the full analysis: Inside My AI Workflow: Building, Digging, And Deciding 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 said on Sept. 29 that he will use Claude Opus 5.5 as his main model for building and the newly released GPT-6.1 Sol for research and review. His workflow update argues that reported cost per task can be as relevant as model scores, while stressing that the benchmark comparison does not settle which system suits an individual workload.

Meyer bases his comparison on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a measure of general capability rather than a verdict on a particular user’s work. In the figures he cites, Opus 5.5 scores 58 at its highest setting and costs $5.98 per task. GPT-6.1 Sol scores 51 at xhigh and costs $0.39 per task. Those figures make Sol a less costly option in this comparison, but the index scores alone do not show how either model will perform on a specific project.

Meyer assigns Opus 5.5 to features, APIs, multi-file work and refactors, using its high setting for regular development and xhigh for difficult work such as architecture or migrations. He assigns Sol to detailed investigation of a file or code change, plus a separate review of Opus’s output. He gives Sol’s high and xhigh costs as $0.32 to $0.39 per task, describing that price as low enough for routine review.

The comparison also covers Claude Sonnet 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6 Luna. Meyer lists them as alternatives for narrower jobs, rather than defaults. He says Sonnet’s high setting is its best value for his uses, while Luna is suited to bulk classification and extraction. These are Meyer’s workflow choices, based on the figures and tasks he describes; the source does not provide independent tests of those assignments.

Why Cost Shapes His Review

Meyer’s main practical point is that a lower model bill can make a second review more affordable, but it does not automatically make the full task cheaper. He says a small token-price saving can be erased by an extra minute of human review. His example is expressly illustrative, not measured, so it should not be read as a general estimate of labor costs.

His approach also makes review a distinct step: one model produces the work, and a model from another family checks it. Meyer says that arrangement is more useful than asking Opus to review its own output. He also cautions that a different model cannot compensate for a flawed specification, and that passing tests alone is not approval to ship. The practical implication is that model selection, review and human judgment all remain part of the workflow.

How the Model Rankings Compare

Meyer frames the market as a price curve: in the index snapshot he cites, six models are within about 20 points, while their listed task costs differ by roughly 100 times. Opus 5.5 leads his table at 58 points; GPT-6.1 Sol is listed at 51 at xhigh, while Luna scores 37 and costs $0.07 per task. The range reflects the particular index and task-cost figures he reports, not a universal price for every user’s prompt.

Effort settings change the comparison. Meyer reports that Opus rises from 51 points and $1.34 per task at medium to 58 points and $5.98 at max. For Sonnet 5.5, he lists max at 56 points and $7.60 per task, compared with 52 points and $2.74 at xhigh. He also reports that max-setting Sonnet generated about 193,000 output tokens per task on the index. These figures explain why his own default is below max, though they do not prove that the same setting is best for other tasks.

GPT-6.1 Sol launched on Sept. 29, Meyer writes, at the same listed token prices as its predecessor: $2 per million input tokens and $10 per million output tokens. Artificial Analysis had published medium, high and xhigh results in the snapshot he used. Meyer notes that the higher settings took 57 to 69 seconds to first token, a delay that may matter for interactive work. He also says the index had not yet published low or max results for Sol.

“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, in the Sept. 29 workflow update

What the Benchmark Cannot Show

The source provides one author’s workflow and a snapshot of third-party index results. It does not report a controlled comparison on Meyer’s own coding tasks, disclose the full task mix behind each cost figure, or establish that the cited results predict performance for other users. Meyer says one index point is inside the noise, and recommends shadow-testing before making a switch.

It is also unclear from the source how the listed task costs translate to a reader’s own usage, including prompt length, output length and selected settings. The source’s final discussion of the cost of human review is incomplete, ending mid-sentence. Its available text therefore does not provide a complete calculation of total task costs.

Test the Workflow on Real Tasks

Meyer recommends shadow-testing models against the work they would handle before changing a workflow. That would let teams compare output quality, latency and cost on their own tasks against their existing process. The source does not give a date for additional benchmark results or report a broader deployment beyond Meyer’s stated choices.

For GPT-6.1 Sol, a next point to watch is whether Artificial Analysis adds the low and max settings that Meyer says were not yet available in the Sept. 29 snapshot. Until then, his recommendation is specific: use Opus for building, Sol for detailed review and research, and test alternatives where the task warrants it.

Key Questions

What changed in Meyer’s workflow?

He says Opus 5.5 remains his main model for building, while newly released GPT-6.1 Sol will handle detailed research and review.

Why does Meyer use GPT-6.1 Sol for reviews?

He cites a task cost of $0.32 to $0.39 at high or xhigh and says that price makes routine review affordable in his workflow. The figures come from the index snapshot he cites.

Does the index show which model is best for every task?

No. Meyer describes the Artificial Analysis Intelligence Index as a measure of general capability, not a verdict on a specific workload, and recommends shadow-testing before switching.

What trade-off does Meyer report for Sol at high settings?

He lists first-token times of 57 seconds at high and 69 seconds at xhigh. That delay could make those settings less suitable for interactive work.

Source: ThorstenMeyerAI.com