How My September 2026 AI Tools Handle Different Jobs

🔍 Read the full analysis: How My September 2026 AI Tools Handle Different Jobs 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’s Sept. 29, 2026, model comparison assigns Claude Opus 5.5 to building software and GPT-6.1 Sol to detailed review, arguing that estimated task costs can differ sharply even among models with relatively close benchmark scores. The report matters to teams choosing AI tools because it frames model selection as a trade-off among quality, effort settings and cost, while its results remain benchmark-based rather than verified for each reader’s work.

The comparison covers six models using Artificial Analysis Intelligence Index v4.3.x scores and the source’s cost-per-task estimates. Opus 5.5 leads at 58 index points in its top setting, with an estimated cost of $5.98 per task. GPT-6.1 Sol reaches 51 at xhigh for $0.39; GPT-6 Luna scores 37 at $0.07. The report lists Sonnet 5.5 at 56 and $7.60, Fable 5.1 at 53 and $7.63, and GPT-6 Astra at 53 and $3.26.

Meyer recommends Opus 5.5 at high for routine development, citing a score of 54 and estimated cost of $1.82 per task. He reserves xhigh, at 56 and $3.46, for work such as architecture and migrations. He places GPT-6.1 Sol at high or xhigh for examining specific files and reviewing changes, and says a different model family can provide a useful second review. These are the author’s workflow choices, not independent evaluations of software quality.

The report says effort settings can change estimated costs substantially. For Opus 5.5, moving from medium to max raises the score from 51 to 58 while increasing estimated task cost from $1.34 to $5.98. For Sonnet 5.5, the listed max setting costs $7.60 for a score of 56, compared with $2.74 and a score of 52 at xhigh. Meyer says this makes Sonnet’s high setting a better fit for scoped tasks and documents.

Cost Shapes the Model Assignment

The report’s central implication is that the highest score may not be the most economical choice for every task. Meyer’s figures put GPT-6.1 Sol’s xhigh estimate at $0.39 per task, compared with $3.26 for Astra and $7.63 for Fable 5.1, while their top scores sit within a few points of one another. If those estimates translate to a team’s own workload, using a lower-cost model for repeated reviews or classification could affect operating costs.

But the comparison itself does not establish that the models produce equivalent results on a particular company’s code, documents or routing decisions. Meyer advises readers to shadow-test models against their own quality bar before switching. He also cautions that savings in model charges may be outweighed by added human review time; his example is illustrative, not a measured result.

A New Model Joins the Comparison

The source dates the releases across September: Fable 5.1 on Sept. 1, Astra on Sept. 3, Opus 5.5 on Sept. 22, Luna on Sept. 22 and Sonnet 5.5 on Sept. 28. GPT-6.1 Sol was released Sept. 29, the day the comparison was published. The author says it has the same listed token prices as GPT-6 Sol: $2 per million input tokens and $10 per million output tokens.

The Artificial Analysis Intelligence Index is described in the source as a measure of general capability, rather than a verdict on a specific workload. The report also lists token output and time-to-first-token figures for Sol: at high, 25 million output tokens on the index and 57 seconds to first token; at xhigh, 36 million and 69 seconds. The author says these longer waits make those settings unsuitable for interactive use.

“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

Workload Results Remain Open

The source does not provide independent, task-specific comparisons showing how the models perform on readers’ workloads. Its per-task costs are estimates tied to the stated index runs, and the material supplied does not fully define the task-cost calculation or its applicability to other usage patterns. Meyer also says one index point is within the noise, and Artificial Analysis had not published low or max settings for GPT-6.1 Sol at the time of writing. The report does not establish whether the listed costs or benchmark standings will hold as models and pricing change.

Test Before Changing Workflows

Meyer’s proposed next step is for teams to shadow-test candidate models on their own tasks before changing assignments. That would let them compare output quality and review time against actual costs. The source gives no date for new benchmark settings or a follow-up comparison, so the timing of further updates is unclear.

Key Questions

Which model does Meyer use for building?

He says he uses Claude Opus 5.5 at high for routine development, and xhigh for harder work such as architecture and migrations.

Why does he use GPT-6.1 Sol for review?

Meyer cites its estimated cost of $0.32 at high and $0.39 at xhigh and uses it to inspect files and diffs. He says a different model family can act as a second reviewer; this is his practice, not a guarantee of independent or correct review.

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

No. The source describes the Artificial Analysis Intelligence Index as a general capability measure and recommends testing models on the work they would actually handle.

What remains unknown about the cost figures?

The supplied report does not fully explain how its per-task estimates translate to other workloads or usage patterns. Teams would need to compare their own quality, usage and human review costs.

Source: ThorstenMeyerAI.com