Opus Builds, Sol Digs, Jev Decides: My September 2026 AI Stack

29 Sept 2026 · Thorsten Meyer

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

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×. That changes the question from “which model is smartest?” to “which model clears my quality bar at the lowest cost per task?”

Here is how I answer it today, on 29 September 2026:

• Opus 5.5 at high or xhigh is my main model for building.
• GPT-6.1 Sol, released today, is the model I use to dig into details and to review.
• Astra, Fable, Sonnet 5.5 and Luna are alternates for specific jobs, not defaults.
• Jev, a decision model that cannot write a sentence, takes over high-volume yes/no and routing judgements.

All scores below come from the Artificial Analysis Intelligence Index v4.3.x unless noted. That index is a map of general capability, not a verdict on your workload, so shadow-test before you switch anything.

Six models, one price curve

Opus 5.5 has the highest score, and GPT-6.1 Sol and Luna are the cheapest ways to get a usable answer. Between them sits a set of models that are hard to justify at their top settings.

Three things stand out. First, Opus 5.5 now outscores its more expensive sibling Fable 5.1 by 5 points and costs less per task. Second, Sonnet 5.5 at max effort costs more per task than Opus at max for 2 fewer points, so it does not belong at that setting. Third, GPT-6.1 Sol costs about one-eighth of Astra and one-twentieth of Fable per task for a score only 1 to 2 points lower.

Prices per 1M tokens (input / output): Opus 5.5 $4 / $20 (cache reads $0.20), Fable and Astra $10 / $50, GPT-6.1 Sol $2 / $10, Luna $0.10 / $0.50.

The effort setting is the real cost lever

Choosing the effort level moves your bill more than choosing between most models. On Opus 5.5, going from xhigh to max adds 2 index points and 73% more cost per task. From medium to max, cost rises 4.46× for 7 points.

That is why I run Opus at high or xhigh for development. High delivers 54 points for $1.82. Xhigh adds 2 more for hard problems such as architecture, migrations and trust boundaries. Max is rarely worth it, and Sonnet 5.5 shows why: at max it writes about 193k output tokens per task, the most Artificial Analysis has measured, and its cost jumps from $2.74 to $7.60 for 4 points.

Medium is still the right default for documents and everyday work. Sonnet’s best value is high, at 47 points for $1.08.

GPT-6.1 Sol: near-Astra scores at a fraction of the price

GPT-6.1 Sol launched today at the same $2 / $10 per 1M tokens as its week-old predecessor, and Artificial Analysis already lists three effort levels. Even the medium setting matches the earlier GPT-6 Sol’s score of 48 at one-fifth of its $1.06 per task.

Compared with the top of the field, Sol xhigh sits 1 to 2 points under Astra (52 at xhigh, 53 at max) and Fable 5.1 (53), at $0.39 instead of $3.26 or $7.63. It is very concise: the high setting used 25M output tokens on the index, against a median of 82M for comparable models.

The catches are real. High and xhigh take 57 to 69 seconds to produce a first token, so this is not an interactive model at those settings. Opus 5.5 still leads it by 5 points at xhigh (56 against 51). Artificial Analysis has not yet published low or max settings, and one index point is inside the noise.

The practical reading: Sol is not the model I ask to build. It is the model I can afford to run on everything, because a review pass at $0.32 to $0.39 per task is cheap enough to be routine.

My stack: Opus builds, Sol digs and reviews

I use Opus 5.5 as my main model and GPT-6.1 Sol as the second pair of eyes. The split follows what each is good at and what each costs.

The review seat matters most. A different model family reviewing Opus’s output is a better check than Opus reviewing itself, and at $0.39 per task I can afford to run it on every meaningful change. Four rules keep that honest:

1. Effort is not capability. Turning up the dial does not make a model smarter.
2. More effort cannot fill in missing requirements.
3. A different model is not an independent review if both read the same flawed spec.
4. Passing tests are not approval to ship.

When something fails review, I hand the failing case and the evidence to Opus, never just “try harder”.

Cheaper tokens are not cheaper work

Halving the model price saves 12.5% of the real cost, and a single extra minute of human review erases it. The example is illustrative, not measured: a task costs $1 of model time plus 4 minutes of review at $45 an hour, so $4.00 in total.

So the number to track is cost per accepted result: model, tools, review and rework, divided by the results someone actually uses. That is the figure that tells you whether GPT-6.1 Sol’s low price saved money or just moved the cost to a person.

Four warnings before you quote any of these scores:

• The index keeps moving. Fable 5.1 scored 66 on an earlier version and 53 on v4.3. Compare only within one version, and treat a 1-point gap as noise.
• Anthropic scores include safety fallback. Flagged cyber and biology tasks route to older models, and that now applies to Sonnet 5.5 too.
• Max effort is not how people run models. Production usually runs medium or high, where gaps narrow and costs fall sharply.
• Your workload decides. Shadow-test on your own tasks and budget cost per task, not cost per token.

Part 2: Jev, the model that decides instead of writing

Jev is the opposite of a frontier model: it cannot write, summarise or extract anything. You send it a piece of text or JSON and a set of typed questions, and it returns a probability and an honest confidence that your code can branch on. It takes about 0.3 to 0.9 seconds per call and costs $0.042 per million input tokens, with output free. At roughly 600 tokens per three-question call, 100,000 decisions cost about $2.50.

There are three question types:

The useful part is the confidence. In my own measurement, Jev agreed with a frontier LLM 97 to 99% of the time when its confidence was 0.8 or higher, and only 42% of the time below 0.5 (89% overall). So the pattern is always the same: act on the clear cases and route the gray zone to something smarter.

Three uses are live in my publishing operation, about 90,000 decisions so far:

I only use Jev when four conditions hold: the volume is high, the question is narrow, a wrong answer is cheap (or unsure cases go to something smarter), and an existing heuristic visibly fails. Then I prove it in shadow first: replay 300 to 500 past decisions, compare accuracy per confidence band, read 20 disagreements, and only wire it in if the high-confidence band reaches 95%. It gets its own flag, off by default, and a canary on 5 to 10 units before rollout.

Know the limits. Jev has no world knowledge, so a bare name needs a snippet of context. It is weaker on maths, dates and non-English text. It reads your wording literally: a rewording shifted my results by about 2 points, so freeze the wording and re-measure after every change. It is early access and hosted only. And it should never be the sole decision-maker for consequences about people, such as hiring, credit, medical or legal outcomes. Jev can sort and flag. A human decides.

What to do this week

Fit beats peak intelligence, in both halves of this article. Route each task to the cheapest model that clears your bar, and re-test when the curve moves. It moved four times in four weeks: Fable and Astra, then Opus, Sol and Luna, then Sonnet 5.5, and now GPT-6.1 Sol.

• Set your main coding model to high or xhigh effort instead of max, and compare cost per accepted result for two weeks.
• Add a review pass with a second model family on every meaningful change. Start with GPT-6.1 Sol at high.
• List your ten highest-volume small decisions (routing, tagging, filtering) and run the four-condition test on each.
• Shadow-test one of them against Jev or Luna on 300 to 500 past decisions before you switch anything on.
• Stop quoting max-effort scores as if they were your production setting.

Sources

Scores and prices are from Artificial Analysis pages opened on 29 September 2026, plus my own earlier reviews. Astra and Fable scores come from the Artificial Analysis Intelligence Index v4.3 announcement and model pages as excerpted in search results.

• GPT-6.1 Sol (medium), GPT-6.1 Sol (high) and GPT-6.1 Sol (xhigh), Artificial Analysis
• Artificial Analysis Intelligence Index overview, including the Claude Sonnet 5.5 coverage
• Announcing Intelligence Index v4.3, Artificial Analysis
• Five frontier models, one bill, Opus 5.5: a new benchmark leader and Opus 5.5: the top model just got cheaper to run
• GPT-6 Sol and Luna: cheaper intelligence, same review bill and OpenAI cuts prices in half
• Which model should write your code?
• Jev and the rise of System One AI
• Jev measurements: my own production data, September 2026, rounded