Sonnet 5 out, Fable 5 back: Anthropic closes June with double-hit
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
The last day of June was a busy one for Anthropic. On June 30, the company released Claude Sonnet 5 — and the same evening received a notice from the U.S. Department of Commerce that it had lifted export controls on Fable 5 and Mythos 5. On July 1, Fable 5 returned to users worldwide after 18 days of forced absence. Let's take both events in order.
Claude Sonnet 5: the most agent-driven mid-range model – with performance close to Opus
What has changed since Sonnet 4? 6
Sonnet 5 promises performance close to the Opus 4.8, but at a significantly lower price. This is a fundamental shift in positioning: Anthropic no longer sees Sonnet as just a cheaper alternative. Sonnet 5 is designed for agency work – tasks where the model must build a plan, use tools, read files, browse pages, execute commands, write or edit code, verify the result and continue without constant user involvement. >
According to Anthropic, Sonnet 5 can make plans, use tools – browsers and terminals – and work autonomously at a level that only a few months ago required larger and more expensive models. >
Specific numbers from benchmarks: on agent programming Sonnet 5 scores 63.2% against 58.1% Sonnet 4.6 and 69.2% Opus 4.8. On the benchmark for working with knowledge Sonnet 5 slightly surpasses Opus 4.8. >>
We gave Claude Sonnet 5 a two-step task – to update Salesforce account levels and send the launch announcement to corporate contacts – and he completed everything from start to finish. It used to get stuck halfway through. For day-to-day automation, it’s an unambiguous choice, said Daniel Shepard, senior engineer at Zapier. >
Cursor’s Sonnet 5 scored 57% on CursorBench versus 49% on Sonnet 4.6 – a notable gain for one of the most used AI development tools. >
Price and availability
Introductory price until August 31, 2026: $2 per million input tokens and $10 per million weekend. After August 31 – $3 and $15, respectively. >
<cite index="5-1" This makes Sonnet 5 cheaper than Opus 4.8, GPT-5.5 from OpenAI and Google’s Gemini 3.1 Pro — albeit more expensive than the Gemini 3.5 Flash. >
From June 30, Sonnet 5 became the default model for Free and Pro plans and is available on all subscriptions, including Max, Team and Enterprise. Also available in Claude Code with a native 1M token context window. >
An important technical detail for developers: Sonnet 5 uses an updated tokenizer that changes text processing to improve performance - similar to changing the tokenizer in Claude Opus 4.7. The same input text can be displayed in 1.0–1.35x more tokens depending on the type of content. The introductory price is set so that the switch to Sonnet 5 is approximately cost neutral. >
Line for API: claude-sonnet-5.
Security
In terms of security, Sonnet 5 shows a lower percentage of “undesirable behavior”—collaborating with misguided requests and deception—compared to its predecessor, making it more secure in agent contexts. It is better at denying malicious requests and evading capture attempts through injections into prompts. Hallucinates and exhibits sycophants less often than Sonnet 4.6. >>
<cite index="10-1" The model is deployed with real-time cybersecurity protection mechanisms similar to those used in Opus models. >
The company is honest about the limitations: the Sonnet 5 is not at the same level as the Opus 4.8 and Claude Mythos Preview when it comes to inconsistent behavior, and its ability to perform dangerous cybersecurity tasks is well below current Opus models. >
Fable 5 is back: chronology of 18 days of lockdown
What happened: recall the chain of events
June 9: Anthropic launched Fable 5 for a wide audience and Mythos 5 for a limited number of trusted cybersecurity partners. Both models are built on the same base, but the Fable 5 comes with more stringent safety barriers. >
June 12: The U.S. government imposed export controls after a report by Amazon researchers showed a way to bypass the security barriers of Fable 5 and force the model to identify software vulnerabilities. In one case, the model created a code showing how one of them could be exploited. >
<cite index="15-1" Because the order went into effect immediately, and Anthropic could not verify users’ citizenship in real time, the company suspended access for everyone – rather than risking a violation of the rule. >
June 13-25: Enterprise users and developers experienced a sudden crash, having to switch to older models such as the Opus 4.8. Tensions reached a peak, with Anthropic publicly objecting, saying the recall of a major commercial model due to a narrow jailbreak finding could "essentially stop all new model releases for all frontline model providers." >
June 26: The government allowed the restoration of access to Mythos 5 for a limited number of trusted American organizations. >
june 30: export controls lifted
The US Department of Commerce has removed export controls on the Claude Fable 5 and Mythos 5 models. It ended the latest dramatic confrontation between the AI company and the Trump administration. >
Commerce Secretary Howard Lutnick notified Anthropic co-founder Tom Brown in a letter about the decision to lift export restrictions, citing Anthropic's close interaction and cooperation with government officials. In the letter, Lutnick said Anthropic agreed to continue cooperating "on protocols, standards and releases for Mythos, Fable and future models," and "inform the U.S. government of any malicious activity." >
Tom Brown led negotiations with the Trump administration, replacing Anthropic CEO Dario Amodei in this role. Amodei has been targeted by the administration for his candid views on AI security and support for Kamala Harris in last election. >
Why it was removed: the model was not a unique threat
<cite index="22-1" Anthropic's own testing showed that competing models were capable of reproducing the same exploit that was the basis for the ban - undermining the government's original rationale. >
The criticism gained momentum, in part because the restrictions handed valuable time to Chinese open-model developers who were trying to catch up with American competitors. >
More than 80 cybersecurity executives have signed a public letter arguing that the opportunity Amazon discovered was not unique to Fable 5.
What has changed technically: a new classifier
Anthropic has developed a new security classifier specifically trained to detect the technology described by Amazon. A specific bypass is now blocked in more than 99% of cases. When the new classifier blocks a request to Fable 5, users receive a notification and the request is redirected to Claude Opus 4.8. >>
Researchers at the Commerce Department’s Center for AI Standards and Innovation tested new and old defenses and called them “exceptionally strong,” although Anthropic acknowledged that a harder filter is also prone to false positives. >
Conditions of return: less than before
For Pro, Max, Team and individual Enterprise subscriptions, Fable 5 is free to use up to 50% of the weekly limit from July 7. After that, access goes to the credit system – that is, the new window is smaller than the original one: six days at a 50% limit versus 13 days of full inclusion before blocking. >
Access to AWS, Google Cloud and Microsoft Foundry will be restored "as soon as possible." >
Mythos 5 is still restricted to about 100 verified American organizations in the Glasswing program. The company says it is continuing to coordinate with the government to expand access. >
What It Means in Context: The New Rules of the Game
June 30 is the day when a new model of relations between American AI laboratories and the state was formed.
<cite index="18-1" Frontier model launches are beginning to look not like regular product releases, but like concerted deployments shaped by U.S. national security oversight — a shift that could slow U.S. spread even as Chinese competitors aggressively move through open channels. >
Anthropic's specific commitments to the government are: to provide government partners with early access to frontline AI models and their defense mechanisms for independent testing prior to public release; prompt notification of government agencies of significant jailbreaks, patterns of abuse and new protective measures; joint research teams; development of industry-wide standards to assess the risks of jailbreaks. >
The company, which spent its early years advocating independence from the military and surveillance purposes, has now agreed to give federal inspectors access to unrealized models before the public sees them - in exchange for the right to sell them. >
Indicative context: locking created a vacuum that was quickly filled by competitors. While Mythos and Fable were offline, OpenAI’s GPT-5.5-Cyber overtook Mythos at CyberGym, a Berkeley benchmark testing AI agents against 1,507 known vulnerabilities. >
What the user gets in the end
On July 2, 2026, the picture is like this.
Sonnet 5 is the new default model for Free and Pro plans, available on all subscriptions and in Claude Code. Agent capabilities previously required by Opus are now available for Sonnet. Until August 31 – at the introductory price of $2/$10 per million tokens.
Fable 5 is back globally on Claude Platform, Claude.ai, Claude Code and Cowork. Until July 7, it is available within the 50% weekly limit on Pro, Max, Team and select Enterprise plans. After July 7, through loans.
Mythos 5 is still only for approved U.S. organizations in the Glasswing program.
For developers, this means a clear practical signal: Sonnet 5 is the model that should be tested first when you need Opus-level refinement of the problem without Opus-level cost. Fable 5 is once again available to those who needed it before June 12 – but with the understanding that it is now not just a product of one company, but a deployment under government control.
*This article is based on official statements by Anthropic, TechCrunch, CNBC, Reuters, VentureBeat, Forbes, Decrypt, Eastern Herald, 9to5Mac, NBC News and Medianama. Actual on July 2, 2026. *
