Anthropic’s Claude Sonnet 5 Hits Opus Coding at Half the Cost

Anthropic launched Claude Sonnet 5 yesterday at an introductory price of $2 per million input tokens and $10 per million output tokens. The rate runs through August 31, 2026, after which standard pricing of $3 per million input tokens and $15 per million output tokens takes effect.

The new model ships with safer defaults than its predecessor, including stronger resistance to prompt-injection attacks and lower rates of hallucination and sycophancy. It also lands inside a charged political moment, the same week the U.S. government lifted export controls on Anthropic’s frontier Claude Fable 5 and Claude Mythos 5. Sonnet 5 cannot, by Anthropic’s own safety design, perform the advanced cybersecurity tasks those two models can, which keeps it under a separate export regime and lets Anthropic promote it to default on every plan including Free.

The New Agentic Mid-Tier Price

Sonnet 5’s headline price is the part of the launch every buyer reads first. Anthropic’s introductory rate is the steepest opening discount the company has offered on a model designed for sustained agentic work. The rate holds through August 31, 2026, after which standard pricing returns to $3 and $15 per million tokens. That standard rate matches the list price Sonnet 4.6 carried at launch.

• Claude Sonnet 5 (intro, through Aug 31, 2026): $2 input / $10 output per million tokens
• Claude Sonnet 5 (standard, from Sep 1, 2026): $3 input / $15 output per million tokens
• Cheaper than: OpenAI GPT-5.5, Google Gemini 3.1 Pro
• More expensive than: Google Gemini 3.5 Flash

Every major lab has now shipped a model that can plan, use tools, and iterate with minimal human input, putting the price of capability rather than the capability itself at the center of the buyer’s decision. Anthropic is now built around a near-flagship agentic model sold at mid-tier prices, a lane the company is staking out across every plan tier. The detailed safety report and benchmark tables sit in the full Sonnet 5 pricing, benchmarks, and safety report on Anthropic’s site.

Anthropic’s new tokenizer maps the same input to between 1.0× and 1.35× more tokens than Sonnet 4.6, depending on content type. The introductory price is set so the transition from Sonnet 4.6 stays roughly cost-neutral for upgrading customers.

Where Sonnet 5 Sits Against the Flagship

Sonnet 5 is positioned to chase flagship performance, not replace it. Anthropic’s launch materials describe the model as a “substantial improvement” over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work. Independent benchmark coverage puts results within striking distance of Opus 4.8 on the agentic tasks that matter most to automation teams.

On agentic coding, Sonnet 5 scored 63.2% on agentic coding against 69.2% for Opus 4.8 and 58.1% for Sonnet 4.6. The same test surfaced Sonnet 5 slightly outperforming Opus 4.8 on a knowledge-work benchmark, the only category where the mid-tier model beat the flagship. Anthropic still recommends Opus 4.8 for the hardest accuracy work. A third-party launch-day breakdown of pricing and partner reactions lays out where Sonnet 5 climbs and where it still gives ground.

Sonnet 5 also gives up access to Anthropic’s most advanced cybersecurity capabilities, by deliberate design. Anthropic reports the model “was never able to develop a full working exploit” in an exploit-development evaluation against Firefox 147, scoring 0.0% on full-success cases. It shows a slightly higher partial-success rate than Sonnet 4.6 on the same evaluation, which Anthropic attributes to general-intelligence gains rather than cyber-specific training.

Sonnet 5 ships with cyber safeguards enabled by default, the same safeguards carried in Claude Opus 4.7 and 4.8. The model is part of Anthropic’s Cyber Verification Program on the native Claude Platform, on AWS, on Azure via Claude in Microsoft Foundry, and on Google Vertex “coming soon.” Organizations enrolled in the program get Sonnet 5 access automatically, per Anthropic’s launch footnote. Anthropic recommends Opus 4.8 for cybersecurity work that requires reduced guardrails. Fable 5, the frontier model released June 9, 2026, has stricter safeguards that block a wider range of cybersecurity tasks and sits under a separate U.S. export regime.

Testers Say the Old Half-Step Problem Is Gone

Anthropic’s launch post carries a stack of partner quotes that describe longer-running workflows finishing on Sonnet 5 where Sonnet 4.6 used to stop. Named testers include Daniel Shepard, a senior engineer at Zapier, and Fabian Hedin, co-founder of Lovable. ClickHouse, Pace, and Eve (a legal-tech firm) appear alongside several unnamed engineering teams. The pattern across the quotes is end-to-end completion: workflows that stalled halfway on earlier models now run through to a tested, verified result.

Anthropic’s testers describe the model tracing failures to root causes instead of patching symptoms, especially on brownfield code with race conditions and hidden tests. At Pace, computer-use agents running insurance workflows on production systems “consistently take the right action and do it quickly.” At Lovable, co-founder Fabian Hedin said in the launch post that the model “refuses unsafe requests cleanly and consistently.”

ClickHouse agents explore live data and produce insights on the fly, so time-to-insight matters when testing new models. Claude Sonnet 5 reasons in tighter steps and gets our users to answers noticeably faster. That speed is a difference our customers feel.

The ClickHouse comment comes from the partner feedback section of Anthropic’s Sonnet 5 launch post. The same end-to-end pattern shows up across the rest of the partner stack, including Zapier on Salesforce automation and Pace on insurance workflows.

Safety Defaults Lift, Cyber Capability Drops

Sonnet 5 ships with safer defaults than Sonnet 4.6 across the categories Anthropic measures. It shows lower rates of hallucination, sycophancy, and other undesirable behaviors, the company reports, and refuses malicious requests more reliably in agentic contexts. It also resists hijack attempts in prompt-injection attacks more often than Sonnet 4.6. On an automated behavioral audit, Sonnet 5 sits below Sonnet 4.6 in misaligned-behavior rate (lower is safer) but above Opus 4.8 and Claude Mythos Preview. Anthropic explicitly recommends Opus 4.8 for cybersecurity work that requires reduced guardrails.

The cyber capability gap is the deliberate part. Anthropic says it did not train Sonnet 5 on cybersecurity tasks, and on a Firefox 147 exploit-development evaluation the model “was never able to develop a full working exploit.” Both Sonnet 5 and Sonnet 4.6 scored 0.0% on full-success cases, with Sonnet 5 showing a slightly higher partial-success rate. Anthropic attributes the partial-success increase to general-intelligence gains rather than cyber-specific training. Anthropic also enrolled Sonnet 5 in the Cyber Verification Program on AWS, Azure, and the native Claude Platform, with Google Vertex access “coming soon,” per the launch footnote.

A Two-Week Washington Standoff Just Lifted

The Sonnet 5 launch landed inside a charged political moment for Anthropic. U.S. export controls had temporarily taken the company’s two frontier models, Fable 5 and Mythos 5, off the global market. The sequence ran as follows:

1. June 9, 2026: Anthropic launches Claude Fable 5 and Claude Mythos 5
2. June 12, 2026: Access suspended after a U.S. export-control directive citing “national security authorities”
3. June 26, 2026: Commerce Secretary Howard Lutnick restores Mythos 5 access to roughly 100 companies and federal agencies
4. June 30, 2026: U.S. export controls fully lifted
5. July 1, 2026: Fable 5 redeploys globally; Sonnet 5 ships to all plans

Anthropic’s earlier statement said the directive ordered the company to suspend all access “by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” After the suspension, a letter from Lutnick to Anthropic co-founder Tom Brown restored Mythos 5 access to a trusted-partner group. Lutnick wrote that the department had “determined that appropriate safeguards are in place” to grant Mythos 5 access. CNBC’s reporting on the deal includes Lutnick’s letter and the list of access recipients.

Fable 5 is back online, with Anthropic proposing a cross-industry jailbreak-severity framework alongside Amazon, Microsoft, Google, and other Glasswing partners. Anthropic’s announcement page carries the original Fable 5 launch post and July 1 global restoration note in one document. Anthropic now runs two strategies in parallel: the unrestricted global shipment of Sonnet 5 on every plan, and a U.S.-anchored flagship rollout still under Washington’s discretion.

An Agentic Race That Now Hinges on Cost

Anthropic is not the only lab repositioning around agentic cost. OpenAI launched GPT-5.6 Sol in preview the week before Sonnet 5, billed as its most agentic model yet. Google shipped Gemini 3.5 Flash in May 2026 as a budget-tier agentic option. Both are pitched as autonomous systems that plan, use tools, and iterate with minimal human input.

The differentiator at every price tier now is how cheaply and reliably an agent runs without human oversight, per TechCrunch’s framing of the launch. Anthropic is selling a near-flagship mid-tier model on the default tier, betting on lane dominance for enterprise automation workloads. Sonnet 5, designed to stay below the export-control threshold, is the model Anthropic can still ship everywhere.

Frequently Asked Questions

How much does Claude Sonnet 5 cost?

Anthropic launched Claude Sonnet 5 on June 30, 2026 at an introductory rate of $2 per million input tokens and $10 per million output tokens. The rate runs through August 31, 2026, after which standard pricing of $3 per million input tokens and $15 per million output tokens takes effect.

How does Sonnet 5 compare to Claude Opus 4.8?

On Anthropic’s published agentic coding benchmark, Sonnet 5 scored 63.2%, against 69.2% for Opus 4.8 and 58.1% for the older Sonnet 4.6. Sonnet 5 slightly outperformed Opus 4.8 on a knowledge-work benchmark. Anthropic still recommends Opus 4.8 for the highest-accuracy tasks.

When did Fable 5 and Mythos 5 come back online?

The U.S. export controls on Anthropic’s Claude Fable 5 and Claude Mythos 5 ended on June 30, 2026, lifting a two-week standoff that began with the suspension on June 12. Fable 5 redeployed globally from July 1, 2026 after Commerce Secretary Howard Lutnick partially restored Mythos 5 access to about 100 companies and federal agencies on June 26.

Is Claude Sonnet 5 on the Free plan?

Yes, Sonnet 5 has been the default model for Free and Pro subscribers since June 30, 2026. It also ships across Max, Team, and Enterprise plans the same day. The introductory $2 / $10 rate applies at every tier until August 31, 2026.

What does the launch mean for OpenAI and Google pricing?

Anthropic’s standard post-promo rate came in cheaper than OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro on launch-day reporting, while Google’s budget-tier Gemini 3.5 Flash remained cheaper. OpenAI’s GPT-5.6 Sol agentic model launched in preview the week before Sonnet 5.