Analysis
Stop picking a model. Pick a mix.
Ask a developer what model they use and you get an honest, messy answer: 'Fable for the hard stuff, Sonnet for the grind, sometimes GPT when I'm out of quota.' Ask the internet and you get a wall of 'Fable 5 vs GPT-5.6' thumbnails. That gap stopped being tolerable this week: GPT-5.6 goes public July 9, and on July 12 Fable 5 leaves Anthropic's paid plans for usage credits. When your best model becomes metered and your competitor's new model goes on sale in the same week, 'which model?' is the wrong question. The right one: which model in which seat?
Quick summary
- On July 7, Anthropic published two internal multi-model patterns: advisor mode (Sonnet 5 executes, Fable 5 advises — ~92% of Fable's SWE-bench Pro score at ~63% of cost, self-reported) and orchestrator mode (~96% at ~46%).
- The structural gap: vendors can't write cross-vendor guidance. Anthropic will only ever put Sonnet in the executor seat; OpenAI will only route you to Sol, Terra, Luna.
- Our arithmetic: with Fable 5 advising, executor swaps land between $0.40 (GLM-5.2) and $1.25 (Sol) per mid-size task, versus $1.80 for Fable solo.
- The honest blank: no one has published quality-retention numbers for cross-vendor pairs. The cost column is checkable; the quality column is empty — that's this week's experiment.
Anthropic just published its own answer — for its own models
On July 7, Anthropic's @ClaudeDevs account published two multi-model patterns the company says it uses internally — the most concrete mixing guidance any frontier lab has shipped, worth taking seriously with sourcing labels attached.
Advisor mode: Sonnet 5 does the work, and Fable 5 is called only at key decision points — Anthropic reports roughly once per task. The reported result: ~92% of Fable 5's standalone SWE-bench Pro score at ~63% of its cost. This ships as a beta 'advisor tool' in the Anthropic API, with specific timing guidance: make the first advisor call early (after a few exploratory reads, before substantive writes) and the final one after file writes and test output exist, before declaring the task done. In other words: buy judgment at the two moments judgment is worth the most — framing the approach, and deciding whether you're actually finished.
Orchestrator mode is the inversion: Fable 5 coordinates, dispatching parallel Sonnet 5 workers via Claude Managed Agents. Reported result on BrowseComp: ~96% of standalone score at ~46% of cost.
To be clear about what these numbers are: all four figures are Anthropic-reported, on Anthropic's benchmarks, and not independently verified. A vendor telling you its expensive model pairs beautifully with its brand-new cheaper model — Sonnet 5 launched weeks ago — is information, not proof. But the shape of the claim matches what mixed-model users have been finding by feel for months: intelligence is worth the most at decision points, and volume work doesn't need the most expensive tokens.
The question Anthropic will never answer
Read the advisor-mode guidance and the next question asks itself: if the executor just needs to be a competent, cheap coding model — why does it have to be Sonnet? GPT-5.6 Terra goes public July 9 at announced pricing in the same band. GLM-5.2 is open-weight under MIT and, per third-party FrontierSWE results, lands near Opus-class on coding at roughly a sixth of Opus's cost.
No vendor will ever walk you through that comparison. Not because they're dishonest — because they're constitutionally incapable of it. Anthropic will only ever put an Anthropic model in the executor seat; it just launched Sonnet 5 and is promoting it, as any company would. OpenAI's guidance will only ever route you through Sol, Terra, and Luna. Every official document about mixing models is also a document about buying more of that vendor's models. The one configuration nobody authoritative will evaluate is the one most users actually want: the best available planner paired with the best-priced adequate executor, regardless of logo.
Cross-vendor guidance can only come from someone with no model to sell. So let's do what we can actually verify — the arithmetic — and be precise about what we can't.
Read the full analysis — free
Sign in to unlock the full Analysis column — no password. You'll also get an alert when big model changes land.
The math, on one honest task shape
We use the advisor-mode shape, since that's the pattern with a published reference point. Assume a realistic mid-size coding task: the executor handles 80,000 input and 20,000 output tokens across all turns (reads, edits, test runs); the advisor — Fable 5, held constant — gets one call of 15,000 in / 2,000 out. At Fable 5's $10/$50 per million, the advisor call costs $0.25 per task, before batch or cache discounts. Now swap executors — everything below is our arithmetic on published or announced prices:
| Executor (in/out per M) | Executor cost | + Fable advisor | Total per task | vs Fable solo |
|---|---|---|---|---|
| Sonnet 5 ($2 / $10)¹ | $0.36 | $0.25 | $0.61 | 34% |
| GPT-5.6 Terra ($2.50 / $15)² | $0.50 | $0.25 | $0.75 | 42% |
| GPT-5.6 Sol ($5 / $30)² | $1.00 | $0.25 | $1.25 | 69% |
| GPT-5.6 Luna ($1 / $6)² | $0.20 | $0.25 | $0.45 | 25% |
| Haiku 4.5 ($1 / $5) | $0.18 | $0.25 | $0.43 | 24% |
| GLM-5.2 (~$0.83 / ~$4.17)³ | ~$0.15 | $0.25 | ~$0.40 | ~22% |
| Reference: Fable 5 does everything | $1.80 | — | $1.80 | 100% |
¹ Promotional pricing through August 31; then $3/$15, which moves the Sonnet row to $0.54 + $0.25 = $0.79 (44%). ² Announced pricing, launching July 9 — verify against OpenAI's published rate card at launch before budgeting on it. ³ Our arithmetic from the circulating "roughly 1/6 of Opus 4.8's $5/$25" figure; approximate. Self-hosted open weights make the real cost compute, not an API meter.
Two honest observations about that table
First: on this task shape, delegation saves more than Anthropic reported — 34% of solo cost with a Sonnet executor versus their ~63%. That's not a contradiction; it means their internal task distribution has a different shape (heavier advisor involvement, different input/output ratios, or simply different tasks). Cost ratios in this pattern depend almost entirely on how much volume the executor absorbs. Your shape will differ from both. The table is a calculator, not a promise.
Second: the executor rows are separated by cents, and the advisor is a fixed toll. Once Fable's $0.25 decision-point cost is locked in, the gap between Sonnet 5 and Terra is fourteen cents per task; between Sonnet and GLM-5.2, about twenty-one. At a thousand tasks a month those cents become real money — but they're small enough that the deciding factor should not be price.
What the math can't tell you
It should be the quality — and that's exactly the number that doesn't exist. Anthropic's ~92% and ~96% retention figures apply to one configuration: Fable 5 advising or orchestrating Sonnet 5. Nobody — not Anthropic, not OpenAI, not any third-party evaluator we can find as of publication — has published equivalent retention numbers for a Fable 5 advisor paired with Terra, Luna, or GLM-5.2. The cost column of our table is checkable arithmetic. The quality column is blank.
There's practical friction too. Anthropic's advisor tool is a beta feature of the Anthropic API — it assumes both seats speak Anthropic. A cross-vendor pair means building your own harness or using tooling that already spans providers, and it means two bills, two rate limits, and two failure modes. None of that is prohibitive for a team that already runs multi-provider infrastructure. All of it is real.
So call this what it is: the experiment the whole industry needs this week. The pattern is published. The prices land July 9. The retention number for cross-vendor pairs is an empty cell that some team with a good eval suite is going to fill first. We'd rather it be filled in public than discovered privately and never shared.
Three mixes for three budgets
- Premium — Fable 5 advisor + GPT-5.6 Sol executor (~$1.25/task on our shape)For work where executor mistakes are expensive and you want frontier judgment at decision points and near-frontier execution. The bet: announced Sol pricing holding at launch, and an unproven cross-vendor pairing outperforming the proven single-vendor one.
- Value — Fable 5 advisor + Sonnet 5 ($0.61, rising to $0.79 in September) or Terra ($0.75). The default.Sonnet is the only executor with a published retention number (Anthropic-reported ~92%) and native advisor-tool support — the least-assumption choice. Terra costs fourteen cents more and bets OpenAI's mid-tier matches Sonnet as an executor: plausible, undemonstrated. If you can run both for a week, you'd be generating exactly the data that doesn't exist yet.
- Floor — Fable 5 advisor, called sparingly, + self-hosted GLM-5.2 (~$0.40/task; near-zero marginal cost self-hosted)For high-volume, lower-stakes work. The bet: your own ops capacity, approximate cost figures, and zero published evidence about how well an open-weight executor follows a Fable advisor's plans. Highest variance, lowest floor.
Whatever you pick: the era of answering 'which model?' with one name ended this week. The vendors each published half a map. The other half — the half that crosses their borders — has to come from somewhere with nothing to sell you.
Need the short answer?
Fable 5 is back worldwide as of July 1 — included in plans through July 12, then usage credits. See the live status, or use GLM-5.2 or the new Sonnet 5 for cheaper work.
Read the brief Fabel 5 spelling guideFAQ
Is advisor mode an official Anthropic feature?
Yes — it ships as a beta 'advisor tool' in the Anthropic API, alongside orchestrator-style workflows on Claude Managed Agents. The 92%/63% and 96%/46% figures are Anthropic's own reported results.
Can I use GPT-5.6 as the executor today?
From July 9, when it goes public. There's no official cross-vendor advisor tool, so you'd wire it yourself or use multi-provider tooling — and accept that no quality-retention data exists yet for that pairing.
Which mix should I start with?
Fable 5 advisor + Sonnet 5 executor — it's the only pairing with a published retention number and native API support. Swap the executor later if your own comparison earns it.
Get the free weekly briefing
One practical email a week: what's worth running on Fable 5, what it costs, and which jobs belong on a cheaper model — plus one email when the next big move lands. No spam, leave anytime.
Sources
This page is independent. Official provider pages are the source of record for access, pricing, and policy.