How Opus, Sol, And Jev Shape My September 2026 AI Stack
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How Opus, Sol, And Jev Shape My September 2026 AI Stack on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get movie nights delivered free — and shop member deals

  • Fast, free delivery on millions of items
  • Access to Prime Big Deal Days deals on October 6–7
  • Prime Video, Amazon Music and more included
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

A Sept. 29, 2026 article describes a personal AI workflow built around Opus 5.5 for development, GPT-6.1 Sol for detailed review and Jev for high-volume decisions. Its benchmark data puts Sol far below several alternatives in reported task cost, but the figures come from one index and the author says teams should test models on their own workloads.

Thorsten Meyer said on Sept. 29 that he uses Claude Opus 5.5 for development, GPT-6.1 Sol for detailed work and independent review, and Jev for high-volume yes-or-no and routing decisions. His account argues that differences in reported task costs now matter as much as small gaps in benchmark scores, but the figures reflect one index and a personal workflow, not a universal ranking.

Meyer bases the comparison largely on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a general capability measure rather than a verdict on any particular workload. In its listed top settings, Opus 5.5 scores 58 and costs $5.98 per task; GPT-6.1 Sol at xhigh scores 51 and costs $0.39. The article lists GPT-6 Luna at 37 and $0.07 per task, with other models between those results.

The workflow assigns Opus 5.5 at high effort to feature work, APIs, refactors and multi-file changes. Meyer says high delivers an index score of 54 at $1.82 per task, while xhigh is reserved for more difficult architecture, migration and trust-boundary work. He reports that xhigh scores 56 at $3.46 per task, and max scores 58 at $5.98.

GPT-6.1 Sol, released Sept. 29 according to the article, is used for examining specific files or diffs and reviewing Opus’s output. At high, the index lists Sol at 50 and $0.32 per task; xhigh scores 51 at $0.39. Meyer assigns Sonnet 5.5 and Luna to scoped side work and classification. He says Jev, which he describes as a decision model that cannot write sentences, handles high-volume routing and yes-or-no judgments; the source provides no benchmark figures for Jev.

At a glance
reportWhen: Published Sept. 29, 2026; GPT-6.1 Sol w…
The developmentThorsten Meyer published a Sept. 29 account of using Opus 5.5, newly released GPT-6.1 Sol and Jev in a cost-based AI workflow.

Opus builds. Sol reviews. Jev decides.

The September 2026 AI stack in one page: six frontier models on one price curve, and a decision model for the high-volume judgements that do not need a sentence.
Scores: Artificial Analysis Intelligence Index v4.3.x. Data as of 29 September 2026.
BuildsClaude Opus 5.5 at high or xhigh effort
Digs and reviewsGPT-6.1 Sol at high or xhigh effort
DecidesJev on high-volume yes/no and routing calls

One price tape, six models

Put every model on the same cost-per-task ruler and capability looks compressed. The bill does not.
Price tape: cost per task of six models on a log scale, from GPT-6 Luna at $0.07 to Fable 5.1 at $7.63$0.05$0.10$0.50$1$5$10cost per task, log scale: each tick is a different order of magnitudeGPT-6 Lunaindex 37 · $0.07GPT-6.1 Solindex 51 · $0.39 (xhigh)GPT-6 Astraindex 53 · $3.26Opus 5.5index 58 · $5.98Sonnet 5.5 · index 56 · $7.60Fable 5.1 · index 53 · $7.63about 100× from the cheapest to the priciest, but only 21 index points between them

Score against cost, at every effort setting

Each dot is an effort level. Opus 5.5 at high already matches Astra and Fable at max on this index, for less money.
Intelligence Index score against cost per task for each effort setting of six models$0.01$0.10$1$102030405060cost per Intelligence Index task, log scaleindexOpus high / xhigh: my defaultOpus 5.5Sonnet 5.5Fable 5.1GPT-6 AstraGPT-6.1 Sol (new)GPT-6 Sol (Sep 22), dashedGPT-6 Lunaup and to the left is better
Astra and Fable are shown at their top published setting. Luna starts at $0.0045 per task. GPT-6.1 Sol has no low or max setting published yet.

The effort dial moves the bill more than the model

Going from medium to max on Opus costs 4.46× more for 7 points. That is why I run high or xhigh.

Claude Opus 5.5

$0.55
42
$1.34
51
$1.82
54
$3.46
56
$5.98
58
low
medium
high
xhigh
max
Solid bars are where I run it. Max adds 2 points over xhigh for 73% more cost.

Claude Sonnet 5.5

$0.41
36
$0.59
41
$1.08
47
$2.74
52
$7.60
56
low
medium
high
xhigh
max
Best value is high. At max it writes about 193k output tokens per task, the most measured.

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

Launched 29 September at $2 in and $10 out per 1M tokens. It sits 1 to 2 points under Astra and Fable, and Opus xhigh still leads it by 5.

Three published settings

SettingIndexCost per taskOutput tokensFirst token
medium48$0.2115M5.3 s
high50$0.3225M57 s
xhigh51$0.3936M69 s
Median for comparable models is 82M output tokens. High and xhigh are not interactive: plan for a wait before the first token.

Same score band, very different bill

GPT-6.1 Sol xhigh
$0.39index 51
Opus 5.5 high
$1.82index 54
GPT-6 Astra max
$3.26index 53
Opus 5.5 xhigh
$3.46index 56
Fable 5.1 max
$7.63index 53
Cost per Intelligence Index task. A one-point gap is inside the noise.

My stack: who builds, who reviews

Opus does the work. A second model family reviews it, because a different reviewer catches what the author cannot see.
Stack diagram: Opus 5.5 builds at high effort, escalates to xhigh, and sends every change to GPT-6.1 Sol for review; Astra or Fable give a second opinionOpus 5.5 · xhighhard problems: architecture,migrations, trust boundariesOpus 5.5 · highMAIN BUILDERfeatures, APIs, multi-filework, refactorsescalate when it gets hardGPT-6.1 Solhigh or xhighdigs into details andreviews every change$0.32–0.39 per taskdifffindingsAstra or Fablesecond opinion, 8 to 20×the cost per taskif they disagreeSonnet 5.5 · Lunaside work: scopedsubtasks, bulk checksand routingFailed review? Hand Opus the failing case and the evidence.Never just “try harder”: effort cannot supply a missing requirement.
Effort is not capability. Turning the dial up does not make a model smarter.
Effort cannot fill gaps. A missing requirement stays missing at any setting.
Different model, same spec. That is not independent review if both read the same flawed brief.
Green tests are not approval. Passing tests only prove what the tests cover.

Cheaper tokens are not cheaper work

Illustrative, not measured: $1 of model time plus 4 minutes of review at $45 an hour. Halving the model price saves 12.5% of the total. One extra minute of review erases it.
$4.00
review $3.00
model $1.00
Baseline
$3.50
review $3.00
model $0.50
Model price cut 50%
$4.25
review $3.75
model $0.50
Cheaper model plus 1 extra minute of review
Track cost per accepted result: model, tools, review and rework, divided by the results someone actually uses.

Read the numbers with four warnings

The index movesFable scored 66 on an earlier version and 53 on v4.3. Compare within one version only.
Fallback is includedFlagged cyber and biology tasks route to older Anthropic models, now on Sonnet 5.5 too.
Max is not productionReal deployments run medium or high, where gaps narrow and costs fall.
Your work decidesShadow-test on your own tasks. Budget cost per task, not per token.

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

Jev cannot write, summarise or extract. It answers narrow typed questions with a probability and an honest confidence, in under a second, for about $0.04 per million input tokens.

One call in, typed answers out

Your code, not Jev, decides what to do with each answer, usually by confidence band.
Jev flow: state and typed questions go into one Jev call; typed answers with confidence come out; code acts alone, escalates the gray zone, or logsStatea ticket, a story,a site profile,a log line …+ typed questions,many per callJevone call0.3 to 0.9 s$0.042 / M tokens inAnswersnoul: 0.03choice: billing p 0.91, conf 0.86score: 2.7 of 3 conf 0.64code branches on thisAct aloneconf ≥ 0.8Escalategray zone toLLM or humanLogmeasure first

Three question types

noul
A yes/no question. Returns the probability of yes, 0 to 1.
gates, flags, filters
choice
Pick one option. Returns the choice, a probability per option, and a confidence.
routing, classification, taxonomy
score
Rate on your ordered levels. Returns a position (it can fall between levels) plus a confidence.
quality, fit, severity, priority

Confidence is the superpower

In my own measurement on a 31-topic classification, Jev agreed with a frontier LLM almost every time it was sure, and rarely when it was not. So: decide the clear cases, route the gray zone.
confidence 0.8 or higher
97–99%
all answers
89%
confidence below 0.5
42%
Agreement with a frontier LLM, my production data, September 2026, rounded.

Three uses running in my publishing operation

About 90,000 decisions so far. Checks I could only afford on a sample now cover everything.
$2.01
Language check
78,889 articles scanned overnight. 1,576 in the wrong language found, 1,553 fixed in place.
22%
Relevance gate
About 10,000 story-to-site pairings judged in 3 days. Only 22% were clearly on-topic.
89%
Classifier fallback
Agreement with the primary LLM across 31 topics, used when that LLM errors.

The fit test, then the shadow test

Use Jev only when all four hold. Then prove it on past decisions before it acts on anything.
High volumeThousands of small calls, not a handful of big ones.
Narrow questionNo multi-step reasoning needed.
Cheap errorsOr unsure cases go to something smarter.
Heuristic failsVisibly, and measured, not assumed.
  1. Replay 300 to 500 past decisions
  2. Compare overall and per confidence band
  3. Read 20 disagreements, decide who was right
  4. High band at 95% or better?
  5. Own flag, off by default
  6. Canary on 5 to 10 units
  7. Roll out in the confident band only

24 use cases, sorted by how well they fit

Start from the strong fits. The amber ones need a measurement before you trust them, and the red ones fail one of the four conditions.
in productionstrong fitmeasure firstpoor 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

No writing, summarising or extractionPair it with an LLM for the write step.
No world knowledgePut a snippet in the state; a bare name means nothing.
Reads your wording literallyA rewording moved my results about 2 points. Freeze it, re-measure after changes.
Weaker on non-English, maths, datesKeep those checks on an LLM. Early access, hosted API only.
100,000 decisions ≈ $2.50
About 60M input tokens at $0.042 per million, output free, roughly 600 tokens per three-question call. Latency 0.3 to 0.9 seconds.
Never the sole decision-maker for consequences about people. Hiring, credit, medical and legal outcomes stay with a human. Jev can sort and flag. A person decides.
Sources. Model scores, cost per task and speeds: Artificial Analysis, Intelligence Index v4.3.x, including the GPT-6.1 Sol medium, high and xhigh pages, checked 29 September 2026. Astra and Fable scores from the Artificial Analysis v4.3 announcement. Jev figures are my own production measurements, September 2026, rounded. The review-bill example is illustrative. Read the full article on thorstenmeyerai.com.

A Workflow Built Around Task Cost

The account illustrates a practical shift in model selection: a small benchmark-score gap may matter less than the price of repeatedly running a model. Meyer says Sol’s review cost is low enough for him to use it on every meaningful change, while keeping Opus on work where he values its higher score. That is his reported practice, not evidence that the same division will suit every team.

His figures also draw attention to the price of effort settings. The article reports that Opus moves from $1.34 per task at medium to $5.98 at max, while its index score rises from 51 to 58. Meyer argues that human review time can erase savings from cheaper model calls, but the supplied source cuts off during its example and gives no measured result for that claim. Readers should treat it as a caution, not a quantified finding.

Benchmarks Behind the Model Split

The source says six models sit within roughly 20 points on the cited index while their listed per-task costs differ substantially. It presents Opus 5.5 as the top scorer among those models, and Sol and Luna as lower-cost options for selected work. Those comparisons depend on the index’s methods and settings; the article cautions readers to shadow-test models before switching.

The author also compares effort levels, which change both scores and costs. For Sonnet 5.5, the source lists high at 47 and $1.08 per task, versus max at 56 and $7.60. It says max produced about 193,000 output tokens per task on the index. These are reported index measurements, not guaranteed costs or performance for other prompts, tools or workloads.

“The index is a map of general capability, not a verdict on your workload, so shadow-test before you switch anything.”

— Thorsten Meyer, in the Sept. 29 article

Limits of the Published Comparison

The source does not specify how the index’s per-task costs map to an individual organization’s actual bills, including its prompts, usage patterns or human review time. It also says the index had not yet published GPT-6.1 Sol’s low or max settings. Meyer notes that a one-point score gap is within the noise, so small differences should not be read as definitive.

Jev is not independently benchmarked in the supplied material, and no test results are given to show how accurately it routes decisions. The source text also ends partway through an illustrative comparison about model prices and human review. It does not provide enough information to establish that example’s outcome.

Test Models Against Your Work

Meyer’s stated next step for readers is to shadow-test models on their own tasks before changing a workflow. That means comparing outputs and costs under the same requirements, then checking where human review time affects the total. The article does not announce a formal follow-up test or a date for updated benchmark figures.

Further comparison will depend on new index results for settings not yet listed and on workload-specific tests. Until those are available, the article’s stack remains one practitioner’s Sept. 29 snapshot, with Jev’s routing role described but not quantified.

Key Questions

What AI stack does Thorsten Meyer describe?

He says he uses Opus 5.5 to build, GPT-6.1 Sol for detailed work and review, and Jev for high-volume yes-or-no and routing decisions. He names Sonnet 5.5, Luna, Astra and Fable as alternatives for selected tasks.

Why does Meyer use GPT-6.1 Sol for review?

He cites its reported cost of $0.32 per task at high and $0.39 at xhigh, which he says makes routine review affordable. Those figures are from the Artificial Analysis index cited in the article.

Does the article establish that Opus 5.5 is best for every developer?

No. It reports Meyer’s own workflow and index comparisons. The article says the index measures general capability and advises readers to test models on their own workloads.

What remains unknown about Jev?

The article describes Jev as a decision model for routing and yes-or-no judgments, but supplies no benchmark scores, task costs or accuracy results for it.

Source: ThorstenMeyerAI.com

FALL

Fall Picks

As an affiliate, we earn on qualifying purchases.

You May Also Like

10 Best OLED Gaming Monitors for Faster, Richer Play in 2026

Discover the best OLED gaming monitors in 2026, featuring top models like Alienware AW3425DW and Samsung Odyssey G5 for immersive, high-speed gaming.

CrowdStrike Outage Impacts Global Microsoft Users

CrowdStrike outage affects Microsoft systems worldwide, disrupting operations and cybersecurity measures for users globally. Stay updated on the impact.

RHEO on Steam: One Toy, Every Screen

RHEO, the fluid art app, is launching on Steam, supporting Windows, Linux, Steam Deck, Steam Machine, and Steam VR, with seamless cloud sync and immersive VR mode.

10 Best Ultrawide Monitors for Work and Gaming in 2026

Discover the best ultrawide monitors for 2026, balancing work and gaming needs. Includes top picks, features, and buying tips.