Opus Builds, Sol Researches, Jev Decides: My AI Workflow

🔍 Read the full analysis: Opus Builds, Sol Researches, Jev Decides: My AI Workflow on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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

The gap between the smartest and cheapest frontier AI models has collapsed into a pricing decision, according to an analysis published on 29 September 2026 by Thorsten Meyer on ThorstenMeyerAI.com. Six leading models now sit within about 20 index points of each other on the Artificial Analysis Intelligence Index v4.3.x, while their cost per task differs by roughly 100x — turning the question from “which model is smartest?” into “which model clears my quality bar at the lowest cost per task?” The same day, OpenAI’s GPT-6.1 Sol launched, and Meyer’s documented response is a workflow in which Claude Opus 5.5 builds, Sol reviews, and a decision model called Jev handles high-volume yes/no judgements.

Meyer’s stack assigns each model a role based on both capability and price. Claude Opus 5.5 (released 22 September, index score 58 at max setting, $5.98 per task) is the main builder, run at high effort (54 points, $1.82 per task) for features and multi-file work, and at xhigh (56 points, $3.46) for architecture, migrations and trust boundaries. GPT-6.1 Sol, released 29 September, handles details and review at $0.32 to $0.39 per task. Sonnet 5.5 (high), Astra, Fable 5.1 and GPT-6 Luna fill scoped roles — agents, subtasks, and bulk classification at $0.07 per task.

Three findings drive the arrangement. First, Opus 5.5 outscores its more expensive sibling Fable 5.1 by 5 points while costing less per task. Second, Sonnet 5.5 at max effort costs more per task than Opus at max for 2 fewer points. Third, Sol xhigh costs roughly one-eighth of Astra and one-twentieth of Fable per task for a score only 1 to 2 points lower. Per 1M tokens, Sol is priced at $2 input / $10 output, against $10 / $50 for both Fable and Astra.

The analysis also identifies effort settings as the dominant cost lever. On Opus 5.5, moving from xhigh to max adds 2 index points for 73% more cost per task; from medium to max, cost rises 4.46x for 7 points. Sol’s trade-off is latency: its high and xhigh settings take 57 to 69 seconds to produce a first token, making it unsuitable for interactive use at those levels, and its output is unusually concise — 25M output tokens on the index at high, against a median of 82M for comparable models, according to Artificial Analysis figures cited by Meyer.

Why Cost Per Task Now Beats Raw Scores

The convergence of scores across model families means capability alone no longer justifies most pricing tiers. If six models clear a user’s quality bar, the decision collapses to economics — and the economics diverge by two orders of magnitude. That reshapes how practitioners buy AI: routine review passes that would be uneconomical at $3 to $7 per task become routine at under $0.40, and cheap models such as Luna ($0.07 per task, 1,429 tasks per $100) make bulk classification and routing viable at scale.

The workflow’s most consequential idea is the cross-family review seat. Meyer argues that a different model family reviewing Opus’s output is a stronger check than Opus reviewing itself, and Sol’s price makes it affordable to run on every meaningful change. The analysis also pushes back on two common assumptions: that effort settings add capability (“turning up the dial does not make a model smarter”) and that model price is the main cost driver — Meyer estimates halving model price saves about 12.5% of real cost, which a single extra minute of human review can erase.

A Month of Back-to-Back Frontier Releases

The six models compared were all released within roughly four weeks: Claude Fable 5.1 (1 September), GPT-6 Astra (3 September), Claude Opus 5.5 (22 September), GPT-6 Luna (22 September), Claude Sonnet 5.5 (28 September) and GPT-6.1 Sol (29 September). GPT-6.1 Sol launched exactly one week after its predecessor, GPT-6 Sol, at the same $2 / $10 per 1M tokens — and even its medium setting matches the earlier model’s index score of 48 at one-fifth the per-task cost, according to Artificial Analysis data cited in the analysis.

All capability figures come from the Artificial Analysis Intelligence Index v4.3.x, which Meyer cautions is a map of general capability rather than a verdict on any specific workload. He recommends shadow-testing — running a candidate model alongside the incumbent on real tasks — before switching.

“In four weeks, the AI frontier stopped being a leaderboard and became a price curve. Six models now sit within about 20 index points of each other, while their cost per task differs by roughly 100x.”

— Thorsten Meyer, ThorstenMeyerAI.com

Where the Data Is Thin or Noisy

Several figures remain incomplete. Artificial Analysis has not yet published low or max settings for GPT-6.1 Sol, so its full cost curve is unknown. Meyer notes that a single index point is inside the noise band, meaning Sol’s 1-to-2-point deficit against Astra and Fable may not be a real difference. The claim that Sonnet 5.5 at max writes about 193k output tokens per task — the most Artificial Analysis has measured — is a vendor-independent measurement but a single data point.

The analysis is also explicitly one practitioner’s workflow, not a benchmark study. The cost-of-ownership example — that halving model price saves 12.5% of real cost — is described by Meyer as illustrative rather than measured. Whether the workflow generalizes to other workloads is untested, and the scores reflect general capability, not performance on any specific task type.

Watch for Sol’s Full Curve and Rival Responses

The immediate developments to watch are Artificial Analysis publishing Sol’s low and max effort settings, which will complete its price-performance picture, and any pricing response from Anthropic or Google to Sol’s sub-$0.40 per-task economics. The release cadence itself — two GPT-6.1 versions within a week, three Claude models in a month — suggests further launches before the price curve stabilizes.

For practitioners, Meyer’s recommended next step is shadow-testing: running a cheaper candidate such as Sol alongside the incumbent model on real workload before committing, and measuring cost per completed task rather than per token. The review seat is the lowest-risk place to start, since a failed review pass costs $0.39 while catching a defect early saves substantially more.

Key Questions

What is GPT-6.1 Sol, and when was it released?

GPT-6.1 Sol is an OpenAI model released on 29 September 2026, one week after GPT-6 Sol, at the same price of $2 / $10 per 1M tokens. According to Artificial Analysis data cited in the analysis, its medium setting matches the earlier model’s index score at one-fifth the per-task cost.

Why use a cheaper, lower-scoring model for review?

At $0.32 to $0.39 per task, review passes are cheap enough to run on every meaningful change. Meyer also argues a different model family provides a better check than a model reviewing its own output, though both models reading the same flawed specification limits that independence.

What does the effort setting actually change?

Per the analysis, effort controls cost and reasoning depth — not capability. On Opus 5.5, going from xhigh to max adds 2 index points for 73% more cost. Sol’s high and xhigh settings carry a 57-to-69-second delay before the first token, ruling out interactive use at those levels.

Are these index scores a reliable guide for choosing a model?

No. Meyer stresses the Artificial Analysis Intelligence Index v4.3.x measures general capability, not performance on a specific workload, and that one index point is within noise. He recommends shadow-testing candidate models on real tasks before switching.

What is Jev’s role in the workflow?

Jev is described as a decision model that cannot write a sentence, used for high-volume yes/no judgements and routing decisions — tasks where generative output is unnecessary and cost per judgement matters most.

Source: ThorstenMeyerAI.com