My AI Stack This September: Opus For Builds, Sol For Digs

🔍 Read the full analysis: My AI Stack This September: Opus For Builds, Sol For Digs on ThorstenMeyerAI.com

Thorsten Meyer says he is using Claude Opus 5.5 for software builds and newly released GPT-6.1 Sol for detailed reviews, citing benchmark scores and task-cost estimates in a comparison published September 29. His figures show substantial cost differences among models with relatively close scores, while leaving open how well the results transfer to individual workloads.

Meyer’s comparison draws on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a measure of general capability rather than a verdict on a specific workload. In the listed top settings, Opus 5.5 scores 58 points at an estimated $5.98 per task. GPT-6.1 Sol at xhigh scores 51 at $0.39. GPT-6 Astra scores 53 at $3.26, while Claude Fable 5.1 scores 53 at $7.63. The article also lists Sonnet 5.5 at 56 points and $7.60 per task, and GPT-6 Luna at 37 points and $0.07.

The source says GPT-6.1 Sol launched September 29 at the same 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 estimated task costs of $0.21, $0.32 and $0.39, respectively. The source reports first-token times of 5.3 seconds at medium, 57 seconds at high and 69 seconds at xhigh. It says the index has not yet published Sol’s low or max settings.

For Opus 5.5, Meyer reports that moving from high to xhigh raises the score from 54 to 56 and the estimated task cost from $1.82 to $3.46. At max, the score reaches 58 and the cost $5.98. He also reports that Sonnet 5.5 at max produces about 193,000 output tokens per task on the index. These are measurements and estimates reported in the source; the material provided does not include a full account of the index’s test method or task-cost calculation.

How Task Costs Shape Model Choice

The comparison makes the cost of a completed task central to model selection. If the reported estimates apply to a team’s own work, a less expensive model could make it practical to run reviews or routine checks more often. Meyer’s proposed division of labor pairs a higher-scoring builder with a lower-cost model from another family for review, rather than using one model for both tasks.

That division is a reported personal workflow, not evidence that every independent model review will catch errors. Meyer himself cautions that a different model can still share the same flawed requirements, and that passing tests alone does not authorize shipping. A review’s value depends on its evidence and criteria as well as its model. His article says teams should shadow-test before switching, a relevant limit because index scores do not establish performance on a reader’s codebase, documents or processes.

The September Model Comparison

The article frames the comparison as a change in how buyers evaluate AI systems: instead of focusing only on which model scores highest, Meyer asks which option meets a quality threshold at the lowest task cost. His table covers models released between September 1 and September 29, 2026, including Fable 5.1, Astra, Opus 5.5, Luna, Sonnet 5.5 and GPT-6.1 Sol. It reports index scores and estimated cost per task, alongside a count of tasks for $100.

According to Meyer’s figures, Opus 5.5 scores five points above Fable 5.1 while costing less per task in the listed settings. He also notes that GPT-6.1 Sol’s xhigh score is one to two points below Astra and Fable, while its estimated task cost is lower. The comparison does not provide uncertainty ranges for these score differences. Meyer cautions that one index point is within the noise, so small gaps should not be treated as decisive rankings.

Effort settings also change the listed costs. Meyer reports that Opus 5.5 rises from $1.34 per task at medium to $5.98 at max, while its score rises from 51 to 58. He recommends high or xhigh for development and medium for everyday documents. Those recommendations describe his approach; the article does not report a controlled test of these settings across different teams or task types.

“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

Where the Benchmarks Leave Gaps

The supplied source does not include a separate evaluation of how these models perform on Meyer’s actual build and review tasks. Its cost figures are presented as estimates per task, but the material provided does not define the task in enough detail to establish that another user would pay the same amount. Index scores measure general capability, Meyer says, and should not be read as direct predictions for a particular workload.

Some results are also incomplete. Artificial Analysis had not published GPT-6.1 Sol’s low or max settings in the comparison, and the source says a one-point index difference falls within measurement noise. It remains unclear how scores and costs might change as more settings are measured, or whether the reported cost ranking holds for specific tasks. The article’s final discussion of human review cost is cut off in the supplied text, so its example cannot be fully assessed.

Testing the Stack on Real Work

Meyer’s stated next step for readers is to shadow-test models on their own tasks before switching. That would let a team compare output quality, time and cost against its current process rather than relying on index scores alone. The source does not give a date for further benchmark updates or say when additional GPT-6.1 Sol effort settings will be published.

For the workflow he describes, Meyer plans to use Opus 5.5 for builds and Sol for detailed reviews, with Astra or Fable as second opinions when the models disagree. Whether that arrangement improves results or reduces total costs remains to be established through use on real work.

Key Questions

What is the main news in Meyer’s comparison?

Meyer says he uses Opus 5.5 for builds and GPT-6.1 Sol for detailed reviews, citing September 29 benchmark scores and estimated costs.

How does GPT-6.1 Sol compare with Opus 5.5?

In the reported top settings, Sol at xhigh scores 51 points at $0.39 per task; Opus 5.5 at max scores 58 at $5.98. The comparison says these are index scores and estimates, not guaranteed results on an individual workload.

Does the article show that Sol is better for every review?

No. Meyer describes his own use of Sol for review. The source does not report controlled results showing that Sol, or any other model, is the best reviewer for every task.

What should teams do before changing models?

Meyer recommends shadow-testing models on their own work before switching. He also cautions that a second model can miss problems if it relies on the same flawed requirements.

Source: ThorstenMeyerAI.com