Sonnet 5 Is the New Default. GPT-5.6 Is Gated. Fable 5 Just Got Expensive. — July 2026 Agent Platform Update

Claude Sonnet 5 at $2/$10 per million tokens — an Opus 4.8-class agent model for Sonnet prices. GPT-5.6 Sol, Terra, and Luna in limited government-gated preview. Fable 5 leaves subscriptions July 7 at $10/$50 per MTok. And the US quietly lifted export controls on Mythos 5. What shipped, what it costs, and what to actually route your agents through.

The last two weeks redrew the agent model map. Anthropic shipped a Sonnet that matches Opus-class performance at a fraction of the price. OpenAI previewed GPT-5.6 in three tiers — but only to 20 government-vetted partners. The Fable 5 subscription window slams shut on July 7, and the US quietly lifted the emergency export controls that killed it three weeks ago.

If you’re routing agent workloads today, your default model just changed. Here’s what shipped, what it costs, and what to actually do about it.

---

Anthropic: Sonnet 5 Resets the Default

Claude Sonnet 5 — June 30

Anthropic released Claude Sonnet 5 on June 30, calling it “the most agentic Sonnet model yet” (Anthropic, June 30). The headline: Sonnet 5 narrows the gap to Opus 4.8 on reasoning, tool use, coding, and computer use — while costing dramatically less.

The cost-performance curves tell the real story. At medium effort, Sonnet 5 is a strict improvement over Sonnet 4.6. At high effort, it matches Opus 4.8 on some tasks — specifically on agentic computer use (OSWorld-Verified) and agentic search (BrowseComp). Anthropic published the curves directly in the announcement: Sonnet 5 at extra-high effort reaches Opus 4.8 territory on both benchmarks.

On knowledge work benchmarks, Sonnet 5 actually edges past Opus 4.8 in some categories (Vellum, June 30). That’s a first for the Sonnet tier.

Pricing — introductory through August 31:

Model	Input / MTok	Output / MTok
Claude Sonnet 5 (intro)	$2	$10
Claude Sonnet 5 (standard, Sep 1+)	$3	$15
Claude Opus 4.8	$5	$25
Claude Fable 5	$10	$50

Sonnet 5 is the new default on Free and Pro plans. It’s available in Claude Code and via the API as claude-sonnet-5. 1M-token context window, 128K max output.

What this means for agent builders: If your agent workload ran on Opus 4.8 for capability reasons, test Sonnet 5 at high effort. At $2/$10 introductory pricing, it’s 60% cheaper on input and 60% cheaper on output than Opus 4.8 — for comparable agentic performance. The ROI math is not subtle.

Early access partners confirmed the agentic leap. Lovable’s Fabian Hedin noted Sonnet 5 “gets more done with less. Same output quality, fewer steps.” A Rust engineer reported Sonnet 5 wrote a reproducing test and fix — then stashed the fix to confirm the bug returned — all in a single unprompted pass.

---

Fable 5: The July 7 Cliff

Fable 5 leaves Claude subscriptions on July 7, 2026 and moves to usage credits at standard API rates: $10 per million input, $50 per million output (Digital Applied, July 1).

The timeline has been chaotic. The original cutoff was June 22. That got pushed to July 7 after the export control debacle. Anthropic says Fable 5 will return to subscriptions “once capacity improves,” but has given no date.

What changes on July 8:

• Pro ($20/mo): Fable 5 moves to usage credits. You get a monthly credit pool (amount TBD by plan) billed at $10/$50 per MTok. Exceed it, and you pay API rates.
• Max / Team / Enterprise: Same model — usage credits replace flat-rate access.

For context: a single agent coding session generating 100K output tokens on Fable 5 costs roughly $5. Under the old subscription, that was effectively free. The economic model for running Fable 5 in autonomous loops changes overnight.

The practical take: Route bulk agent work to Sonnet 5. Reserve Fable 5 for the hardest problems where Mythos-class reasoning actually changes the outcome. We covered the underlying TCO dynamics in our enterprise agent cost breakdown — the math hasn’t gotten friendlier.

---

Export Controls: Lifted, Quietly

On June 30, the US Department of Commerce lifted the emergency export ban on Claude Fable 5 and Mythos 5 that was imposed June 12 (CNBC, June 30). Anthropic immediately redeployed both models globally (The Guardian, July 1).

The ban lasted 18 days. In that window, Anthropic cut off international API access, disabled the models for non-US users, and sparked a legal fight that reached the courts. Commerce Secretary Lutnick’s June 26 letter created an “Annex A” list of roughly 100 approved US entities — critical infrastructure defenders, national labs, and government agencies — who retained access during the blackout.

Mythos 5 remains restricted to that Annex A list even after the general lift. Fable 5 is globally available again — but only through July 7 on subscriptions, and at full API rates after that.

The longer-term signal: export controls on frontier models are now a live operational risk, not a policy hypothetical. If your agent infrastructure depends on a single model provider’s top tier, you need a routing fallback. We covered the architectural implications in our export control analysis from June.

---

OpenAI: GPT-5.6 Preview — Powerful but Gated

OpenAI previewed GPT-5.6 on June 26 in three tiers: Sol (flagship), Terra (balanced), and Luna (fastest/cheapest) (OpenAI, June 26). The catch: only about 20 government-approved partners have access. Broader availability is promised “in the coming weeks.”

Verified Benchmarks

On Terminal-Bench 2.1, the numbers are significant (QCode.cc):

Model	Terminal-Bench 2.1
GPT-5.6 Sol Ultra	91.9%
GPT-5.6 Sol	88.8%
Claude Mythos 5	88.0%
GPT-5.6 Terra	84.3%
GPT-5.5	83.4%
GPT-5.6 Luna	82.5%
Claude Opus 4.8	78.9%

Sol Ultra at 91.9% is the highest publicly reported Terminal-Bench score. Terra at 84.3% matches GPT-5.5-class performance at roughly half the price of Sol.

Pricing (per million tokens)

Tier	Input	Output
GPT-5.6 Sol	$5	$30
GPT-5.6 Terra	$2.50	$15
GPT-5.6 Luna	$1	$6

Cached reads keep a 90% discount. The context window hasn’t been officially disclosed — rumors point to 1.5M tokens, but OpenAI hasn’t confirmed.

The government gate is new. The staged rollout is at the US government’s explicit request, tied to GPT-5.6’s elevated cybersecurity capabilities. OpenAI’s system card confirms Sol and Terra “can find vulnerabilities and pieces of exploits” but stop short of the “Critical” risk tier (OpenAI Deployment Safety Hub).

For builders: GPT-5.6 Terra at $2.50/$15 is priced directly against Sonnet 5 at $2/$10 (introductory). The benchmark gap between Terra (84.3%) and what Sonnet 5 can do at high effort (matching Opus 4.8 at 78.9%) makes this a genuine competition — once both are generally available. Right now, only one of them is.

---

The Broader Landscape

Agentic AI Foundation Goes Live

The Linux Foundation launched the Agentic AI Foundation (AAIF) with founding contributions from Anthropic (MCP), OpenAI (AGENTS.md), and Block (goose) (Linux Foundation). MCP is now under neutral governance — a significant shift for the protocol that’s become the de facto standard for agent-tool communication. We’ve covered MCP’s role in the agent protocol stack in depth.

IPO Pipeline

Anthropic’s confidential S-1 (filed June 1) and OpenAI’s (filed ~June 8) are both in SEC review. Anthropic disclosed $47B annual revenue at a $965B post-money valuation, with $1.25B/month going to SpaceX for compute (Reuters). Neither company is GAAP-profitable. The public market window is open — but the cash-burn numbers are staggering.

EU AI Act: ~28 Days to Enforcement

August 2, 2026. If your AI products have EU users and you haven’t completed a risk classification exercise, the window is now under a month. Fines reach €35M or 7% of global turnover.

---

What to Actually Do

If you’re routing agent workloads today:

1. Default to Sonnet 5. At $2/$10 introductory pricing through August 31, it’s the best cost-performance ratio for agentic workloads. Test at high effort if your task needs Opus 4.8-class capability.

2. Plan for the Fable 5 cliff. After July 7, every Fable 5 call costs real money. Reserve it for tasks where Mythos-class reasoning demonstrably changes outcomes. Route everything else to Sonnet 5.

3. Watch the GPT-5.6 rollout. When Terra hits general availability, it’ll compete directly with Sonnet 5 on price and capability. Have an API key ready. If you’re building multi-model routing, GPT-5.6 Terra belongs in your eval matrix the day it ships.

4. Don’t bet your infrastructure on a single frontier model. The export control saga proved that model availability can change overnight. Multi-provider routing isn’t optional anymore — it’s infrastructure hygiene.

The model tier explosion we predicted in June’s platform update is accelerating. Sonnet 5 proves the trend: mid-tier models are absorbing capabilities that required flagship models three months ago. The economic pressure on $10/$50 per-MTok pricing isn’t theoretical — it’s here, and it’s coming from below.