Anthropic Scraps Claude Sonnet 5 Price Hike in Enterprise Pivot

Anthropic has scrapped a planned 50% price increase for Claude Sonnet 5, making its introductory API rate of $2 per million input tokens and $10 per million output tokens permanent on August 10. The decision cancels the $3 per million input and $15 per million output standard pricing that was scheduled to take effect on September 1. Analysts at ValueAddVC described the move as a concession to competitive pressure in the frontier-model price war, where providers have been cutting token costs to win enterprise workloads.

The August 2026 update cycle marks a strategic pivot for Anthropic. Instead of emphasizing raw model capability, the company is now competing on enterprise economics and governance: stable pricing, self-hosted infrastructure, security scanning, compliance APIs, and workforce training. For executives and founders, the signal is that agentic AI is becoming procurement-ready, with pricing stability, auditability, and deployment control as the new battleground.

Pricing reversal and enterprise economics

Anthropic’s August 10 announcement confirmed that the introductory rate for Claude Sonnet 5 would not expire. The standard rate of $3 per million input tokens and $15 per million output tokens had been set for September 1, but the company reversed course before that date. The permanent rate of $2 per million input tokens and $10 per million output tokens removes a near-term cost increase for developers and enterprises that had been planning around the higher standard pricing. Analysts at ValueAddVC read the decision as a direct response to competitive pressure in the frontier-model price war, in which rival providers have been undercutting each other on token costs to attract enterprise workloads.

The pricing decision is not isolated. It sits alongside a broader enterprise push: self-hosted Claude Code environments, third-party skill scanning, and general availability of compliance-related APIs. Together these moves suggest Anthropic is trying to remove procurement objections rather than simply tout benchmark scores. The company appears to be betting that predictable unit economics and deployment controls will matter more to enterprise buyers than incremental gains on public benchmarks.

Enterprise controls: self-hosting, scanning, and protocol maturity

On August 6, Anthropic opened self-hosted Claude Code environments to public beta for Team and Enterprise plans. Sessions run on customer infrastructure, although prompts and tool results still go to Anthropic for inference. The feature is disabled by default and is not available to organizations with zero-data-retention requirements. This means customers gain more control over where code execution happens, but they do not yet get full data isolation from Anthropic’s inference services.

The same day, Anthropic launched skill and plugin security scanning in beta for Enterprise customers. The system performs automatic pass/warn/fail checks on third-party skills and plugins, with scans typically completing in one to two minutes. On August 20, computer use, the browser tool, the Skills API, and the Files API reached general availability. These releases expand the surface area for enterprise integration, but the security scanning still excludes MCP servers, hooks, and organizations using customer-managed encryption keys, zero-data retention, or HIPAA configurations.

Underlying these changes is the MCP 2026-07-28 specification, the protocol’s largest revision since launch. It moves to a stateless core, strengthens OAuth/OIDC, and introduces versioned extensions such as Apps and Tasks, with a 12-month deprecation window for older specs. The protocol update is designed to make integrations more auditable and easier to secure, aligning with the enterprise governance theme of the month.

Safety, transparency, and the biology safeguard retune

On August 7, Anthropic retuned the biology safeguards for Claude Fable 5. Testing showed roughly 85% fewer biology-related fallbacks, with total fallbacks down 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the API. Dual-use topics — including virology, toxicology, and molecular design — still fall back to Opus 5, a less capable model for those requests. The retune reduces unnecessary refusals for benign biology queries, but it does not unlock professional research use cases that touch dual-use areas.

On August 17, Anthropic’s Transparency Hub published model-report summaries for Opus 5, Sonnet 5, Fable 5, and Mythos 5. Two days later, the Opus 5 system card changelog reported that attacker success on the Gray Swan IPI benchmark fell from 5.5% to 2.0% at 15 attempts compared with Opus 4.8 — a 98% resistance rate. The improvement suggests that safety interventions are becoming more robust against repeated attack attempts, though the benchmark measures a specific class of indirect prompt injection.

Despite these improvements, Fable 5 remains effectively unusable for professional biology research and drug development, because dual-use requests are rerouted to the less capable Opus 5. That limitation is a continuing source of criticism from researchers who need frontier-level reasoning on virology, toxicology, or molecular design tasks.

Product expansion, limits, and unresolved questions

August also brought a series of product and training updates. On August 14, “Auto mode” became the default for new Claude Code sessions on Pro, Max, and Team plans. On August 18, Claude Cowork fully rolled out to mobile and web for all paid plans. On August 20, Anthropic launched Claude Academy, offering free courses, tutorials, badges, and a “4D AI Fluency” framework based on the company’s employee onboarding. On August 25, memory was extended to Claude Cowork with editable topics and a sensitive-topics opt-in; it is on by default for Free, Pro, and Max plans, and off for Team and Enterprise plans.

These moves point to a deliberate effort to make agentic tools easier to adopt and govern inside organizations, not just for individual developers. But several caveats remain. Skill scanning excludes MCP servers, hooks, and organizations using customer-managed encryption keys, zero-data retention, or HIPAA configurations. Anthropic states that a “pass” is explicitly not a safety guarantee. Self-hosted Claude Code shifts operational burden to customers and still sends conversation data to Anthropic for inference, which may limit its appeal for highly regulated environments.

“A ‘pass’ is explicitly not a safety guarantee.”

It is unclear whether Anthropic will extend the permanent Sonnet 5 pricing to other models or whether the biology safeguard retune will eventually allow Fable 5 to serve professional research use cases. The company has not said when or if self-hosted sessions will support zero-data-retention deployments.

Looking ahead, the August 2026 updates suggest Anthropic is betting that enterprise adoption will be won through governance and operational controls as much as through model performance. If the permanent Sonnet 5 pricing holds and the compliance tooling matures, the company may be able to convert its safety reputation into procurement advantage. The open question is whether rivals will respond with further price cuts or faster feature parity — and whether Anthropic can close the remaining gaps in self-hosting, security scanning, and dual-use research without sacrificing its safety commitments.