Opus Builds And Sol Digs—Here’s How Jev Decides

Why Cost Per Task Now Drives Model Choice

The six-model cluster within 20 index points means raw capability differences have narrowed while price differences have not. For anyone running AI at volume, the operative question becomes which model clears a given quality bar at the lowest cost per task, not which model tops a leaderboard. Sol’s $0.39-per-task xhigh setting makes routine second-pass review economically viable — Meyer notes a review pass at $0.32 to $0.39 per task is cheap enough to run on every meaningful change.

The review seat matters most in this framing. A model from a different family reviewing Opus’s output is a stronger check than Opus reviewing itself, Meyer argues, because it avoids shared blind spots — though he cautions that a different model is not an independent review if both work from the same flawed specification.

Meyer also warns that cheap tokens do not equal cheap work. Halving model price saves only about 12.5% of real cost in his illustrative example, and a single extra minute of human review can erase the saving — a reminder that model spend is one component of total workflow cost.

How the September Model Wave Shaped Up

The analysis caps a four-week stretch in which six frontier-tier models shipped between 1 and 29 September 2026, according to Meyer’s timeline: Fable 5.1 on 1 September, GPT-6 Astra on 3 September, Opus 5.5 and GPT-6 Luna on 22 September, Sonnet 5.5 on 28 September, and GPT-6.1 Sol on 29 September. Sol launched at the same published price as its week-old predecessor — $2 input and $10 output per 1M tokens — but its measured per-task costs come in far lower, partly because of how concise it is: Meyer reports Sol’s high setting used 25M output tokens on the index, against a median of 82M for comparable models.

Even Sol’s medium setting matches the earlier GPT-6 Sol’s score of 48 at one-fifth of that model’s $1.06 per-task cost. Token prices across the compared models range from Luna’s $0.10/$0.50 per 1M input/output to $10/$50 for Fable and Astra, with Opus 5.5 at $4/$20 and cache reads at $0.20.

“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 100×.”

— Thorsten Meyer, ThorstenMeyerAI.com

Caveats in the Sol Comparison

Artificial Analysis has not yet published low or max settings for GPT-6.1 Sol, so its full performance range is unknown. Meyer also notes that one index point is inside the noise, meaning the 1-2 point gaps between Sol and higher-priced models may not reflect real capability differences on any specific workload. The index itself measures general capability, not performance on a particular user’s tasks — Meyer explicitly recommends shadow-testing before switching models. His cost figures beyond model spend, including the 12.5% saving from halved prices, are described as illustrative rather than measured.

Waiting on Sol’s Full Benchmarks

The immediate open item is Artificial Analysis publishing Sol’s low and max effort settings, which would complete the comparison table and clarify whether the model scales further at the top end. Meyer’s stated workflow — Opus 5.5 at high or xhigh for building, Sol at high or xhigh for review, Jev for high-volume yes/no and routing judgments, and Sonnet 5.5 high plus Luna for side work — will be tested against real workloads rather than index scores. Readers considering any similar arrangement should shadow-test candidate models on their own tasks before committing, since benchmark proximity does not guarantee equivalent results on a specific workload.

Key Questions

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

GPT-6.1 Sol is an OpenAI-lineage model released on 29 September 2026 at $2 input and $10 output per 1M tokens. According to Artificial Analysis figures cited by Thorsten Meyer, its xhigh setting scores 51 on the Intelligence Index v4.3.x at $0.39 per task.

Why pair Claude Opus 5.5 with GPT-6.1 Sol instead of using one model for everything?

Meyer’s reasoning is economic and practical: Opus 5.5 scores highest (58 at max, 56 at xhigh) for building work, while Sol delivers 51 points at $0.39 per task, making it affordable to run as a routine second-pass reviewer. A model from a different family also provides a check that self-review cannot.

What is Jev, and what role does it play?

Jev is a decision model that Meyer says cannot write a sentence but handles high-volume yes/no and routing judgments. It occupies the cheap, high-volume slot in his stack where generative output is unnecessary.

Does a higher effort setting make a model smarter?

No, according to Meyer: ‘Effort is not capability.’ Higher effort settings raise cost substantially — on Opus 5.5, medium to max multiplies cost 4.46× for 7 index points — and cannot compensate for missing requirements.

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

Not on their own. Meyer describes the Artificial Analysis index as a map of general capability rather than a verdict on any specific workload, and notes that single index points fall within measurement noise. He recommends shadow-testing any model before switching.

Source: ThorstenMeyerAI.com