GPT-5.6, Fable 5, Grok 4.5: A Marketing Routing Matrix

Multi-model routing for marketing teams stopped being optional this week. GPT-5.6 — the Sol, Terra, and Luna family — went GA across ChatGPT, Codex, and the OpenAI API today, July 9, 2026. Grok 4.5 shipped yesterday. Anthropic’s Fable 5 and Sonnet 5 pair, a month and nine days into market respectively, is the incumbent every new release gets measured against.

Here’s the fact that should reframe how you buy AI capacity: on OpenAI’s own GA benchmark tables, Claude models lead four of the headline evals — SWE-Bench Pro, GDPval-AA Elo, the AA Intelligence Index, and HealthBench Professional — while GPT-5.6 Sol leads the AA Coding Agent Index. The lab that published the tables didn’t win them all. If OpenAI’s own launch page argues against a single crowned champion, your model strategy should too.

This post is the dated cross-vendor snapshot no single lab publishes: the July 9 price sheet across all three vendors, recomputed cost math per finished marketing task, and a task-to-model routing matrix for five concrete marketing jobs — from high-volume tagging to client strategy memos. The engineering playbook for routers, fallbacks, and eval harnesses already covers the architecture; this is the “which model, which task, what cost” layer that sits on top of it.

The state of the lineup, as of this morning. OpenAI took GPT-5.6 GA today across ChatGPT, Codex, and the API — “available starting today,” with full availability promised over the next 24 hours. The family ships as three tiers: Sol (frontier), Terra (mid), and Luna (volume), with a reasoning-effort ladder that now runs from none through a new max setting. Our GPT-5.6 GA breakdown covers the tier family, ultra mode, and access by plan in full — we won’t re-derive it here.

Yesterday, SpaceXAI shipped Grok 4.5 — a mixture-of-experts model trained jointly with Cursor on tens of thousands of NVIDIA GB300 GPUs in its Memphis data centers, served at a stated 80 tokens per second with a 500K context window. And Anthropic’s current pair — Fable 5 (launched June 9) and Sonnet 5 (June 30, default on Free/Pro plans since July 1) — is the reference point both launches priced and benchmarked against.

Why does a same-day, three-lab snapshot matter? Because each vendor’s launch material compares against a strategically chosen subset of rivals, at a strategically chosen moment. Nobody publishes the whole board on the same date with the same cost math. The rest of this post does exactly that — explicitly dated, so you can treat it as a snapshot to re-verify, not an evergreen ranking to trust indefinitely.

The most useful benchmark evidence this week comes from an unusual place: the launching vendor’s own tables. OpenAI’s GA page shows Claude Fable 5 ahead of GPT-5.6 Sol on GDPval-AA v2 Elo (1,759.6 vs 1,747.8), the AA Intelligence Index v4.1 (59.9 vs 58.9), and HealthBench Professional (60.9% vs 60.5%) — while Sol leads the AA Coding Agent Index v1.1 (80 vs 77.2). On SWE-Bench Pro, the software-engineering eval most relevant to teams building their own marketing tooling, the picture looks like this:

The interpretation matters more than any single score. When the lab publishing the benchmark table shows a competitor leading four of its headline evals, the era of “pick the best model and use it for everything” is over — what remains is task-level comparative advantage. Fable 5 looks strongest where sustained judgment and multi-step engineering are the job; Sol looks strongest where agentic coding throughput is the job; Grok 4.5 concedes the accuracy crown and competes on tokens, speed, and price. For the fuller head-to-head on the previous generation’s frontier, see the full Grok 4.5 vs Opus 4.8 vs GPT-5.5 benchmark breakdown.

What follows is our own composite of three separately sourced vendor pricing tables — OpenAI’s GA page, x.ai’s developer pricing docs, and Anthropic’s platform pricing docs (the Claude rates re-verified live against platform.claude.com today). No vendor publishes this comparison; each individual rate is that vendor’s own list price.

Ranked by output price per 1M tokens, cheapest first: Claude Haiku 4.5 at $5, then GPT-5.6 Luna and Grok 4.5 tied at $6, Claude Sonnet 5 at $10 (introductory, through August 31), GPT-5.6 Terra at $15, Claude Opus 4.8 at $25, GPT-5.6 Sol at $30, and Claude Fable 5 at $50. That’s a 10× spread from the cheapest lane to the priciest — which is precisely why routing exists as a discipline. Input rates follow the same shape: $1 (Haiku, Luna) up to $10 (Fable 5).

Two pricing mechanics deserve a line each. First, caching: OpenAI bills cache writes at 1.25× the uncached input rate with a 90% discount on reads and a 30-minute minimum cache life; Anthropic charges Fable 5 cache reads at $1 per 1M tokens; Grok 4.5 prices cached input at $0.50 per 1M. For prompt-heavy marketing workloads — long brand guidelines, style guides, product catalogs re-sent on every call — cache economics can matter more than the headline rate. Second, the earlier Sol-vs-Fable-5 price and access comparison from July 2 still holds at GA — Sol’s list price is half of Fable 5’s at both ends, unchanged from preview.

List prices only become decisions when you multiply them by a realistic workload. Take the most common volume job in a marketing operation: classification — lead scoring, content categorization, email triage. Assume a typical call consumes 1,000 input tokens and returns 200 output tokens. Ten thousand such calls burn 10M input and 2M output tokens. Every derived cell below recomputes from that formula against each model’s list rate — cost = 10 × input rate + 2 × output rate.

Read the endpoints: the identical 10,000-call job costs $20 on Haiku 4.5 and $200 on Fable 5. If a tagging pipeline is running on a frontier model out of habit, that habit costs 10× at list rates — before quality even enters the conversation, because tagging rarely needs frontier judgment.

But the table also shows where sticker math stops being enough. Grok 4.5’s stated 4.2× output-token reduction versus Opus 4.8 (max) on SWE-Bench Pro means its effective cost on agentic work is lower than its $32 row suggests — a model that finishes the task in a quarter of the tokens bills a quarter of the output. The same logic applies to GPT-5.6’s Programmatic Tool Calling numbers, with a critical caveat: those 24–63.5% reductions came from customers adopting PTC as an engineering pattern, not from switching a model slug. Cost per finished task — tokens actually consumed to reach an accepted output — is the metric to instrument, and it requires measuring your own pipelines rather than reading any vendor’s page, including this one.

Five concrete marketing job types, one recommended model each, with the price and the sourced reason. This is our synthesis — no vendor publishes a cross-lab task matrix — and it deliberately routes by comparative advantage on the evidence above, not by lab loyalty. Prices are list rates per 1M tokens (input / output) as of July 8–9, 2026.

Three notes on reading the matrix. In the volume lanes, the three cheapest models sit within $12 of each other per 10,000 calls — so the tiebreakers are operational: existing SDK integration, caching fit, data-handling terms. In the judgment lanes, the split recommendation on agentic coding is deliberate: Sol and Fable 5 each lead a relevant vendor benchmark, and the honest answer is to run both against your own repos before defaulting. And in the office lane, Grok Build’s native Excel, PowerPoint, and Word automation is genuinely differentiated — Danny Wu, Canva’s Head of AI Products, credits GPT-5.6 with strong slide-creation evals, but document automation as a product surface is where x.ai planted its flag. For how Grok 4.5 stacks up on creative and agentic work more broadly, see how Grok 4.5 compares on creative and agentic work. If content production at scale is the lane you care most about, that’s the operating model behind our content engine service — routing volume drafting to volume-priced models and reserving senior review (human and frontier-model) for judgment passes.

Four non-benchmark facts materially change the matrix for specific teams — and none of them appear in headline comparisons.

Data retention on Fable 5. Anthropic’s Mythos-class models — which include Fable 5 — carry mandatory 30-day data retention, and Zero Data Retention agreements do not cover them. For an agency routing client-confidential strategy drafts or unreleased campaign material, that’s a routing input in itself: some client engagements may require the judgment lane to run on a model whose retention terms the client has signed off.

Sonnet 5’s intro price expires. The $2 / $10 rate holds through August 31, 2026; from September 1 it rises to $3 / $15 — a 50% increase on both ends that would take our 10,000-call example from $40 to $60. Anthropic frames the intro price as “roughly cost-neutral” rather than a discount, because Sonnet 5 and Fable 5 use a newer tokenizer producing approximately 30% more tokens for the same text — that’s Anthropic’s own framing, but the tokenizer effect is real budget math either way: comparing $/1M across tokenizers understates Claude costs slightly at equal text volume.

Ultra mode is not a cost lever. GPT-5.6’s ultra mode runs four agents in parallel by default (some GA charts show 16-agent configurations), trading higher token spend for stronger results and faster wall-clock time. It belongs in the quality-escalation column of a routing policy, never the cost-savings column. And a migration note from OpenAI’s own docs worth carrying into any cutover plan: treat migration as a tuning pass, not only a model-slug change — GPT-5.6 is reportedly more sensitive to “be concise” instructions than prior generations.

A matrix on a page is a policy; production routing is an engineering system. We’ve deliberately kept this post to the “state of the lineup” layer because the build layer is already covered: the engineering playbook for routers, fallbacks, and eval harnesses covers how to route, fall back, and evaluate, and our team-and-budget stack decision tree covers which stack fits which team. What this snapshot adds is the July 9 answer to the variable those guides parameterize: which model goes in each slot today.

The forward trajectory seems clear from this week alone: the labs are converging on list prices in the volume tier — three models within a dollar of each other on output — and differentiating on efficiency multipliers instead. Grok 4.5 markets token frugality; OpenAI markets PTC-driven token cuts; Anthropic reprices through a tokenizer change. Expect the next competitive round to be fought on effective cost per finished task, which means routing decisions will increasingly need your own telemetry — tokens consumed per accepted output, per task type, per model — rather than any vendor’s benchmark table. Three revisit triggers are already on the calendar: Grok 4.5’s expected EU availability in mid-July, GPT-5.6’s unpublished context window and rate limits whenever OpenAI ships them, and Sonnet 5’s September 1 price step. If you want help instrumenting that — or standing up the routing layer itself — this is the core of our AI transformation engagements.