🔍 Read the full analysis: The AI Workflow I’m Using This September, From Opus To Jev on ThorstenMeyerAI.com
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
TL;DR
In a Sept. 29 account, Thorsten Meyer describes using Opus 5.5 for building, GPT-6.1 Sol for detailed review, and Jev for high-volume yes-or-no and routing decisions. His model and cost comparisons draw on Artificial Analysis Intelligence Index v4.3.x, which he says should be treated as a general capability guide rather than a measure of performance on every workload.
Thorsten Meyer says he is using Claude Opus 5.5 to build, GPT-6.1 Sol to examine details and review changes, and Jev for high-volume yes-or-no and routing judgments. In a Sept. 29 account, he frames the workflow around differences in task cost among models whose Artificial Analysis Intelligence Index scores are relatively close, while cautioning that the index does not establish which model will perform best on a particular workload.
Meyer assigns Opus 5.5 at high effort to features, APIs, multi-file work and refactoring, citing an index score of 54 and a listed cost of $1.82 per task. For architecture, migrations and trust boundaries, he uses the xhigh setting, which scores 56 at $3.46 per task in the supplied table. He says he rarely uses max effort: Opus rises from 56 at xhigh to 58 at max, while the listed cost rises from $3.46 to $5.98.
GPT-6.1 Sol, released Sept. 29 according to the source, is his second model for file-level investigation and review. Its listed high and xhigh costs are $0.32 and $0.39 per task, with index scores of 50 and 51. Meyer says those settings can take 57 to 69 seconds to produce a first token, making them less suited to interactive use. The source also assigns scoped subtasks and documents to Sonnet 5.5, routine classification to Luna, and occasional second opinions to Astra or Fable.
The account identifies Jev as a decision model that cannot write sentences. Meyer uses it for high-volume binary decisions and routing. The source provides no benchmark score, price, release date or detailed evaluation method for Jev, so its role is described as part of his workflow rather than a quantified comparison with the other models.
Opus builds. Sol reviews. Jev decides.
One price tape, six models
Score against cost, at every effort setting
The effort dial moves the bill more than the model
Claude Opus 5.5
Claude Sonnet 5.5
GPT-6.1 Sol: near-Astra scores at a fraction of the price
Three published settings
| Setting | Index | Cost per task | Output tokens | First token |
|---|---|---|---|---|
| medium | 48 | $0.21 | 15M | 5.3 s |
| high | 50 | $0.32 | 25M | 57 s |
| xhigh | 51 | $0.39 | 36M | 69 s |
Same score band, very different bill
My stack: who builds, who reviews
Cheaper tokens are not cheaper work
Read the numbers with four warnings
Part 2: Jev, the model that decides instead of writing
One call in, typed answers out
Three question types
Confidence is the superpower
Three uses running in my publishing operation
The fit test, then the shadow test
- Replay 300 to 500 past decisions
- Compare overall and per confidence band
- Read 20 disagreements, decide who was right
- High band at 95% or better?
- Own flag, off by default
- Canary on 5 to 10 units
- Roll out in the confident band only
24 use cases, sorted by how well they fit
Proven in production
- 1Relevance gate
- 2Language check
- 3Classifier fallback
Publishing and content
- 4Thin-source detector
- 5Same-event dedupe
- 6Product fits roundup
- 7Disclosure present
- 8Headline quality
- 9Comment moderation
Commerce and support
- 10Support-ticket routing
- 11Return-reason coding
- 12Review to feature complaints
- 13Catalogue taxonomy
- 14Order-fraud pre-triage
Software and AI systems
- 15LLM guardrail
- 16RAG passage filter
- 17Citation check
- 18Tool and intent routing
- 19Log-line triage
- 20PR risk triage
Business ops and home
- 21Inbox triage
- 22Expense categorisation
- 23Lead qualification
- 24Smart-home intent
Limits, cost and one hard rule
How the Workflow Splits Model Costs
The account illustrates a practical way to use several AI models in development: reserve a more expensive model for work that calls for building or difficult reasoning, then use a cheaper model for repeatable review and routing. Meyer’s stated rationale is that a $0.39 review task can be run routinely, while a second model family may catch issues the builder missed. That is his operating judgment, not evidence that this arrangement improves software quality in every team or project.
The figures also show why the effort setting matters to his choices. In the supplied index table, Opus at max costs more per task than at xhigh for a two-point score difference. Sonnet 5.5 at max is listed at $7.60 per task for a score of 56; Meyer notes that its xhigh setting is $2.74 for a score of 52. These are index task costs, not a promise of what an individual user will pay. Actual costs depend on the work submitted and the usage behind it.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Bose SoundLink Max Portable Bluetooth Speaker, IP67 Waterproof, Black
As an affiliate, we earn on qualifying purchases.
The Index Behind Meyer’s Choices
Meyer bases the listed scores on the Artificial Analysis Intelligence Index v4.3.x, except where otherwise noted. In his account, Opus 5.5 scores 58 at its top setting; GPT-6.1 Sol scores 51 at xhigh; Astra scores 53 at max; and Fable 5.1 scores 53. Luna scores 37. He describes six models as sitting within about 20 index points while their costs per task differ substantially. The figures and rankings here reflect the supplied Sept. 29 source and may change as benchmark listings are updated.
His accompanying comparison lists token prices as input and output rates per million tokens: Opus 5.5 at $4/$20, Fable and Astra at $10/$50, GPT-6.1 Sol at $2/$10, and Luna at $0.10/$0.50. Those rates are distinct from the source’s per-task cost estimates. Meyer also says the index is a general capability map, not a verdict on a reader’s own work, and recommends shadow-testing before switching workflows.
““which model clears my quality bar at the lowest cost per task?””
— Thorsten Meyer
What the Benchmarks Cannot Establish
The supplied material does not report results from a controlled comparison of these models on Meyer’s own projects. It does not give the prompts, task mix or calculation method behind every per-task estimate, and it supplies no comparable performance or cost data for Jev. The figures should therefore be read as the source’s account of benchmark results and Meyer’s workflow, not as independently verified predictions for other users.
Meyer says a one-point index difference is within the noise and that Artificial Analysis had not yet published low or max settings for GPT-6.1 Sol. He also reports long first-token waits at Sol’s high and xhigh settings. The account does not establish whether those measurements will hold across different tasks or change as the model and index are updated.
Test the Workflow on Real Tasks
Meyer’s next step for readers is to shadow-test models against their current workflow before switching. That means checking candidate outputs against the same quality bar and tasks used in practice. The source does not give a timetable or results from a broader test, so it remains unclear whether the proposed split will generalize beyond his own use.
For review failures, Meyer says he sends the failing case and supporting evidence back to Opus rather than simply asking it to try harder. His account also cautions that passing tests alone is not approval to ship and that different models can still share the same flawed requirements. No further release or benchmark milestone is specified.
Key Questions
Which model does Meyer use to build?
He uses Opus 5.5 at high effort for features, APIs, multi-file work and refactoring, and xhigh for harder architecture and migration tasks.
What role does GPT-6.1 Sol play?
Meyer uses GPT-6.1 Sol for detailed investigation and independent review. The supplied account lists its high and xhigh task costs at $0.32 and $0.39.
What is Jev used for?
The source describes Jev as a decision model that cannot write sentences. Meyer assigns it high-volume yes-or-no decisions and routing judgments; no benchmark or price is provided.
Do the index scores show which model is best for every task?
No. Meyer describes the Artificial Analysis index as a general capability guide and says readers should test models on their own workloads before switching.
Source: ThorstenMeyerAI.com
