Claude Haiku 5.5 Explained: 75% Cheaper, 1M Token Context

Anthropic released Claude Haiku 5.5 on October 7, 2026, calling it “the cheapest, fastest, and most capable small model we’ve ever released.” The model is built for high-volume, cost-sensitive work: summaries, document compaction, database queries, classification requests, live customer support, browser use, and as a subagent running alongside Opus 5.5 and Sonnet 5.5 on coding tasks.

The headline number is cost. Anthropic says Haiku 5.5 is roughly 75% cheaper to run on average than Haiku 4.5, while also extending the context window to 1 million tokens and adding an adjustable effort setting that previous Haiku models did not have.

Here is what Anthropic has confirmed, what it costs, and what changes for anyone building on the Claude API.

What Anthropic Announced

On October 7, 2026, Anthropic released Claude Haiku 5.5, the newest entry in its small-model line. In its own announcement, the company described it as “the cheapest, fastest, and most capable small model we’ve ever released.” The model is positioned for high-volume, cost-sensitive work rather than frontier reasoning: summaries, context compaction, database queries, classification, live customer support, browser use, and running as a subagent alongside Opus 5.5 and Sonnet 5.5 on coding tasks.

Anthropic also used the same announcement to cut the cache-read price of Claude Sonnet 5.5 in half, from $0.20 to $0.10 per million tokens, which the company says reduces Sonnet 5.5’s cost on most agentic work by about 20%. A new monthly API credit for Claude Max and Team subscribers was announced alongside both changes.

What's Confirmed

Every figure in this article comes from Anthropic's own announcement at anthropic.com/news and the model card at anthropic.com/claude-haiku-5-5, cross-checked against the Claude Platform documentation at platform.claude.com/docs/en/models/overview. Nothing here is a rumor or a leak; it is Anthropic describing its own release.

• Context window: 1,000,000 tokens, up from 200,000 tokens on Haiku 4.5.
• Max output: 128,000 tokens, up from 64,000 tokens.
• Knowledge cutoff: Anthropic lists June 2026 as the model's reliable cutoff.
• Effort levels: Haiku 5.5 is the first Haiku-class model with an adjustable effort setting: Low, Medium, High, Xhigh, and Max. Medium is the default. Thinking is adaptive and on by default, and can be turned off at the Low, Medium, or High effort levels.
• Model ID: claude-haiku-5-5, available now on the Claude Platform, AWS, Google Cloud, and Microsoft Azure.

Why the Price Cut Matters

The practical effect of a 75% average cost reduction is that workloads which were previously too expensive to run on every request, high-frequency classification, large-scale document summarization, or an always-on customer support bot, become cheap enough to run constantly. That shift also makes the subagent pattern more affordable: a company can let Opus 5.5 or Sonnet 5.5 act as the lead model on a coding task while delegating narrow lookups to Haiku 5.5 without the subagent calls adding meaningfully to the bill.

Cognition's Walden Yan described exactly this pattern in Anthropic's announcement, using Haiku 5.5 as a sidekick model in Devin Fusion alongside Opus 5.5 as the lead, a combination that held a top-tier FrontierCode score of 66.2. Rogo's Alex Wang said the company uses Haiku 5.5 as a subagent for pulling specific data points, such as 10-K segment revenue, out of longer documents.

Who's Affected

The announcement is aimed at developers building on the Claude API and enterprises already running Claude in production, not at the average person using claude.ai through the chat interface. Haiku-class models already sit behind parts of the free tier of claude.ai, so most consumer users will not notice an immediate change. The companies that do notice are the ones with high request volumes: customer support platforms, CRM tools, document-analysis products, and coding agents that lean on a cheap subagent for routine lookups.

Benchmark Performance

Anthropic published benchmark comparisons against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 (shown for reference as the tier above Haiku):

The gap over Haiku 4.5 is largest on agentic and computer-use tasks, which lines up with Anthropic's positioning of Haiku 5.5 as a model meant to handle multi-step, tool-using work rather than just one-shot text generation.

What Changes for Developers

Switching to Haiku 5.5 means updating the model ID in API calls to claude-haiku-5-5. The effort parameter is new: Medium is the default, and raising it to High, Xhigh, or Max trades latency and cost for more thorough reasoning on harder subtasks, while Low favors speed. Anthropic notes that prompts up to 100,000 tokens cover roughly 90% of requests that previously went to Haiku 4.5, which is why the pricing below is split at that threshold.

What It Costs

Pricing is per 1 million tokens, split by prompts up to 100K tokens and prompts over 100K tokens, compared with Haiku 4.5 and Sonnet 5.5:

Sonnet 5.5's cache-read price was also cut in half as part of the same announcement, from $0.20 to $0.10 per million tokens, which Anthropic says lowers Sonnet 5.5's cost on most agentic work by around 20%.

Where It's Available

Claude Haiku 5.5 is available now through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure, in the same regions Anthropic already supports for its other current models. Anthropic has not announced India-specific pricing or availability differences, this is a global API and cloud-platform release rather than a region-specific one.

Early Customer Results

Anthropic's announcement included results from several companies already running Haiku 5.5 in production. Asana's Aaron Vinh, a staff software engineer, reported a 30%-plus latency reduction and up to 2.5x faster inference per agent turn. HubSpot's Ze'ev Klapow said the model posted the company's best-ever score on its CRM evaluation suite, at 92.8%. AlphaSense's Daniel Campos measured 0.84 versus 0.76 over Haiku 4.5 on document question-answering. Box's Yashodha Bhavnani saw an 11-point higher score at roughly half the latency compared with Haiku 4.5.

Safety Changes

Anthropic reports major alignment improvements over Haiku 4.5 in its own evaluations, including fewer instances of misaligned behavior and lower willingness to cooperate with misuse. Cybersecurity safeguards are more restrictive than Haiku 4.5's but less restrictive than Sonnet 5.5 or Opus-class models; the model still blocks penetration testing and attacker-style techniques. Biology safeguards sit at the same tier as Sonnet 5, Sonnet 5.5, and Opus 5.

What's Next

Anthropic has not announced a specific roadmap item tied to Haiku 5.5 beyond this release. The company has not said whether a dedicated claude.ai consumer update is planned, so for now the model's reach is primarily through the API and the three cloud platforms listed above.

Frequently asked questions

Related articles