How Three AIs Fit Into My September 2026 Workflow

🔍 Read the full analysis: How Three AIs Fit Into My September 2026 Workflow 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 uses Claude Opus 5.5 to build and the newly released GPT-6.1 Sol to investigate details and review work, in a September 29 account of his AI workflow. The split reflects a reported gap between model scores and costs: Meyer cites benchmark results placing several models close together in capability, while their cost per task varies widely.

Meyer names Opus 5.5 at high or xhigh effort as his main model for features, APIs, multi-file work and refactoring. He uses xhigh for harder problems such as architecture, migrations and trust boundaries. He assigns GPT-6.1 Sol at high or xhigh to focused investigations of files or diffs and to a second review pass. The account describes this as a division of work, not a claim that one model performs best on every task.

The figures Meyer cites come primarily from the Artificial Analysis Intelligence Index v4.3.x. In that index, Opus 5.5 scores 54 at high effort and 56 at xhigh, with cited costs of $1.82 and $3.46 per task. Sol scores 50 at high and 51 at xhigh, at $0.32 and $0.39 per task. Meyer says the index measures general capability and does not decide which model suits a particular workload.

Other models have narrower roles in his account. Sonnet 5.5 and Luna are options for scoped subtasks, documents and bulk classification; Astra or Fable may serve as another opinion when Sol and Opus disagree. Meyer also describes Jev as a decision model for high-volume yes-or-no and routing judgments. The source excerpt gives no benchmark table or cost figures for Jev, so its performance cannot be compared here.

Why Meyer Splits Building and Review

The workflow shows how reported cost per task can shape model selection even when benchmark scores are relatively close. Meyer cites Sol xhigh at 51 index points and $0.39 per task, compared with Astra at 53 points and $3.26, and Fable 5.1 at 53 points and $7.63. These are figures from the source’s cited index and task-cost estimates, not a guarantee that Sol will deliver comparable results on every user’s work.

Using a model from a different family for review is meant to provide a second perspective on the first model’s output. Meyer says that review is affordable enough to run routinely at the stated Sol costs. He also cautions that a second model cannot repair missing requirements in the original specification, and that passing tests alone is not approval to ship. Those limits matter for teams weighing automated review against human judgment.

Effort settings also affect the comparison. Meyer reports that moving Opus 5.5 from xhigh to max raises the index score by two points and the cost per task by 73 percent. That supports his choice of high or xhigh for development, but the figures describe the index’s measured tasks; the source does not establish that the same trade-off applies to every project.

The September Model Cost Comparison

Meyer frames the shift as a move from choosing a model by leaderboard position to considering quality at a given cost per task. His table lists six models within about 21 index points at their listed top settings: Opus 5.5 at 58, Sonnet 5.5 at 56, Fable 5.1 and GPT-6 Astra at 53, GPT-6.1 Sol at 51, and GPT-6 Luna at 37. These scores are attributed to Artificial Analysis Intelligence Index v4.3.x, with the listed release dates spanning September 1 to September 29.

The source also compares token prices, which are distinct from its per-task estimates. It lists 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. Those rates do not by themselves establish the cost of a particular job: the source’s per-task amounts depend on benchmark task usage, and should not be treated as a universal bill for similar work.

Meyer reports that Sol launched on September 29 at the same listed token rates as its predecessor, GPT-6 Sol. He says Sol’s medium index setting scored 48 at a cited cost of $0.21 per task, while high and xhigh scored 50 and 51. The source notes that one index point is within the noise and that the index had not yet published Sol’s low or max settings. Those gaps limit what can be inferred from the early comparison.

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

— Thorsten Meyer, in the September 29 workflow account

What the Benchmark Cannot Show

The account reports index scores and task-cost estimates, but does not provide enough detail to independently reproduce the per-task costs or determine how closely benchmark tasks match Meyer’s actual projects. It also does not establish that the same ranking holds across different prompts, software stacks or quality requirements. The source itself advises shadow-testing before a switch.

Sol’s high and xhigh settings are reported to take 57 and 69 seconds to produce a first token in the index. Meyer says those settings are therefore not suited to interactive use, but the excerpt does not explain the measurement method or how response times vary across real workloads. The available material also ends during a section on the human cost of review, so its full argument and any supporting figures for that section are unavailable.

It remains unclear how often Meyer uses the alternate models, how he resolves disagreements between reviewers, and whether the cited costs include human review or other operational expenses. The source says that one index point is within the noise; small score differences should not be read as decisive on the evidence provided.

Shadow-Test Before Switching

Meyer recommends shadow-testing models against the existing workflow before replacing a default. That would let a team compare outputs on its own tasks and judge whether lower reported costs meet its quality bar. The source gives no timetable, test protocol or results from such a comparison, so it does not establish what another team should expect.

For Meyer’s own approach, the next step described is continued use of Opus for building and Sol for detail work and review, with alternatives reserved for jobs where tests support them or where the two models disagree. The source does not state whether he plans to change these assignments as more index settings or results become available.

Key Questions

What is the main development in Meyer’s account?

Meyer describes a workflow split in which Opus 5.5 handles most building and GPT-6.1 Sol handles focused investigation and review. He published the account on September 29, 2026, the same day he says Sol was released.

Why does Meyer use GPT-6.1 Sol for review?

He cites a reported cost of $0.32 to $0.39 per task for Sol at high or xhigh and says a different model family provides another perspective on Opus’s output. That is his rationale; the source does not establish that this method catches every kind of error.

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

No. Meyer says the Artificial Analysis Intelligence Index measures general capability and is not a verdict on an individual’s workload. He recommends shadow-testing before switching models.

What remains unknown about the comparison?

The source does not provide a reproducible cost methodology or show how the benchmark results translate to other users’ tasks. It also says some Sol effort settings had not yet been published and that a one-point score difference is within the noise.

Source: ThorstenMeyerAI.com