Editor’s note (Aug 22, 2026): This edition was revised after publication. The version published Aug 18 described a three-tier lineup with Opus 4.8 as the workhorse and missed two launches that had already shipped — Sonnet 5 (June 30) and Opus 5 (July 24). It also predated Anthropic’s August 2 content-marking rollout. The tracker below is rebuilt on the current lineup; the export-control analysis is unchanged from the original.

Four models in seven weeks, an eighteen-day disappearance, and a compliance change that touches every email your team drafts. Q3 was the busiest quarter Anthropic has had, and the headline is not any single launch — it is that the frontier tier stopped being the interesting one. This is our quarterly read on what actually changed, what it costs, and which workflows justify moving.

The Q3 2026 lineup

The working set for GTM builders is now four tiers, and three of them are new since our Q2 edition.

Model	Shipped	API list (per M in / out)	What it is for
Fable 5	June 9	$10 / $50	Frontier tier; genuinely hard reasoning and very long autonomous runs
Opus 5	July 24	$5 / $25	The new default for judgment-dense agentic work
Sonnet 5	June 30	$2 / $10	Mid-tier agents; planning, tool use, browser and terminal work
Haiku 4.5	(carried over)	$1 / $5	Volume classification, routing, first-pass enrichment

Two footnotes. Mythos 5 — the same frontier model with some safeguards lifted — exists but is restricted to approved research partners, so it is not a GTM consideration. And Fable 5 moved to usage-credit access on Claude subscription plans after June 22, so team-level budgeting matters in a way it did not when Opus-class models were the default.

Opus 5 is the quarter’s real story

Fable 5 got the announcement, but Opus 5 is the model that changes architectures. It shipped at $5 / $25 — the same price as Opus 4.8 — while Anthropic describes it as coming “close to the frontier intelligence of Claude Fable 5 at half the price,” with performance that “more than doubles” on its Frontier-Bench evaluation versus its predecessor. Both are vendor claims and belong in your own evals before they change anything, but the pricing is a fact, and the pricing is the point: the same budget now buys materially more capability at the tier most GTM agents actually run on.

The capability that matters most for revenue work is not raw intelligence anyway — it is self-correction over long chains. Anthropic’s own framing is that Opus 5 “was much stronger at verifying its work and iterating carefully until it succeeds.” That is precisely the failure mode that breaks GTM agents in production: not a wrong answer on step one, but an unnoticed wrong answer on step seven that the agent confidently builds on for another twenty steps. If that holds up on your workloads, it is worth more than any benchmark delta.

Sonnet 5 deserves more attention than it got. At $2 / $10 — pricing Anthropic has confirmed is now permanent rather than introductory — it is described as “the most agentic Sonnet model yet,” able to “make plans, use tools like browsers and terminals, and run autonomously,” with performance “close to that of Opus 4.8.” Read that carefully: the previous workhorse tier’s capability is now available at 40% of its price. Any agent you built on Opus 4.8 in Q2 deserves a Sonnet 5 A/B before you renew your budget.

The June export-control interruption — and why it belongs in your vendor risk file

On June 12, a US export-control order forced Anthropic to suspend global access to Fable 5 after a reported safeguard bypass; the order was lifted June 30 and the model returned worldwide on July 1 with a retrained classifier. The episode is unprecedented and its lesson is operational, not political: frontier-model availability is now a regulatory variable. If your enrichment waterfall, research agents, or outbound drafting run on a single frontier model with no fallback path, an eighteen-day suspension is an outage you did not architect for. Teams that pinned workflows to one model spent June rewriting under pressure; teams with model-agnostic routing, per our agent-native playbook, changed a config value.

New in August: everything Claude writes now carries a mark

This is the development most likely to reach your team this quarter, and it arrived after this edition first published. From August 2, 2026, models launched on or after that date embed an imperceptible watermark directly into generated text, with signed C2PA provenance metadata attached to supported files. The trigger is EU AI Act Article 50, which became enforceable on that date — but Anthropic is applying the marking worldwide, not just in the EU. Models released before August 2 fall under a transition period, with coverage targeted before December 2, 2026.

What this means in practice, from Anthropic’s own documentation: you will not see it, it does not change the meaning or readability of the output, it persists when text is copied, and it “may persist through some editing.” Heavy editing, format conversion, or unsupported platforms can degrade it.

Three implications for GTM teams, in descending order of how much they should change your behaviour:

• Your AI-drafted outbound is now potentially attributable. Not to you — to Claude. A detected mark means the content “may have been processed by Claude,” which is a much weaker statement than “this was AI-written.” It cannot distinguish a fully generated cold email from a human-written one Claude proofread.
• Absence of a mark proves nothing. If you were hoping to use detection to screen inbound content or vet an agency’s work, note that marks are lost through heavy editing and never existed on other vendors’ output. Anthropic has promised detection tooling and technical documentation but has not shipped a public detector yet.
• It is a compliance asset more than a liability. If your legal or procurement function has been asking how you would evidence AI involvement in customer-facing content, this answers it — and it does so on the vendor’s side, at no cost to you.

Our read: this is not a reason to switch models, and treating it as one would be an overreaction. It is a reason to stop pretending AI-drafted outreach is indistinguishable from human-written, and to make sure your disclosure posture is a deliberate choice rather than an accident.

The fallback design, and why GTM workflows will not hit it

Fable 5 ships with safety classifiers that silently route certain requests — offensive cybersecurity, biology and chemistry, and model-distillation attempts — to a lower tier rather than refusing them. Anthropic reported at launch that more than 95% of sessions involve no fallback at all, and the restricted categories are essentially orthogonal to revenue work: enrichment, drafting, scoring and call analysis run on the full model. One caveat worth checking rather than assuming: the published fallback target was Opus 4.8, and with Opus 5 now shipping at the same price point that routing may well have moved. Verify before you depend on it.

Practical upgrade checklist: which model for which workflow

Move to Opus 5 now: anything you are running on Opus 4.8. Same price, materially better, and it is the tier where long-horizon self-correction pays for itself — deep account research, pre-call synthesis across large document sets, CRM-wide analyses, RevOps migrations.

Test Sonnet 5 against Opus 4.8 workloads: if an agent worked acceptably on the old workhorse tier, Sonnet 5 at $2 / $10 may hold quality at a fraction of the cost. This is the single highest-ROI experiment available to a GTM team this quarter.

Reserve Fable 5 for genuinely frontier-hard tasks: multi-hundred-account territory analysis, research runs measured in hours, work where you have tested Opus 5 and it demonstrably falls short. At $10 / $50 it is the wrong default.

Keep Haiku on volume: classification, routing, first-pass enrichment. And price-check it — OpenAI’s July 30 cut put GPT-5.6 Luna at $0.20 / $1.20, well under Haiku’s $1 / $5, which is a real consideration for high-volume, low-judgment work. We work through that trade in the rebuilt Claude vs GPT comparison.

The pattern from our Build-vs-Buy analysis holds and has only sharpened: spend frontier tokens on judgment, commodity tokens on everything else. There are now four tiers to apply that rule across.

What has not changed

Tool-use reliability, computer use, and Agent Skills — the step changes we documented last quarter — carry forward. MCP remains the integration spine, now on the finalized 2026-07-28 stateless spec with enterprise-managed authorization promoted to stable. And the fundamental tracker rule stands: model launches are vendor claims until your own evals say otherwise. Run yours on your own accounts before re-platforming anything.

Forward look: Q4 2026

Watch four things. Third-party benchmark coverage of sustained-autonomy tasks — the capability that matters most for GTM agents and is still the least independently measured. Whether the export-control episode repeats; one more suspension makes multi-model routing a hard requirement rather than good hygiene. Price movement at the mid-tier, where Sonnet 5 and GPT-5.6 Terra are now matched at $2 input and the competition is on output economics. And the December 2 watermarking deadline for pre-August models, plus whether a public detection tool actually ships — because a marking scheme nobody can verify is a compliance artifact, not a transparency measure. We will re-baseline in the Q4 edition.