![Anthropic Launches Claude Haiku 5.5, Cuts Prices 75% [2026]](https://tech-insider.org/wp-content/uploads/2026/10/tmp5dh_9ok0.webp)
Anthropic shipped Claude Haiku 5.5 on October 7, 2026, and the headline number is the one that matters most to engineering teams watching their API bills: input tokens under 100,000 now cost $0.10 per million, with output priced at $0.50 per million. For prompts that cross the 100,000-token mark, pricing steps up to $0.50 input and $2.50 output per million tokens. Anthropic, in its official release notes, says that works out to roughly 75% cheaper to run than Haiku 4.5 on average, with the company’s own breakdown putting the real savings at 90% for prompts up to 100,000 tokens and 50% for anything longer, a cut that lands squarely in the middle of an API pricing war that has been running across every major AI lab since late 2025. The launch was first flagged for a wider audience by outlets including Unite.AI, PYMNTS, and CyberSecurityNews.
The model is live immediately through Anthropic’s own platform, plus Amazon Web Services, Google Cloud, and Microsoft Azure, according to Anthropic’s release notes. That’s the third model in the Claude 5.5 line to ship in roughly a month, following Claude Opus 5.5 and Claude Sonnet 5.5. Anthropic is calling Haiku 5.5 its “cheapest, fastest, and most capable small model,” and the pitch is aimed squarely at the high-volume, cost-sensitive workloads that make up the bulk of production AI traffic: classification, summarization, data extraction, customer support, voice agents, and in-app assistants.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What’s actually new in Claude Haiku 5.5
The context window is the most concrete spec upgrade. Haiku 5.5 ships with a 1 million-token context window, a fivefold jump from the 200,000-token ceiling on Haiku 4.5. That difference matters more than it sounds, because context window limits are usually what force developers to upgrade from a cheap model to a pricier one mid-pipeline. A support-ticket classifier that used to choke on a long thread history, or a document-extraction job that had to chunk a 300-page PDF into fragments, can now run on the budget model end to end.
Anthropic also reports meaningful speed gains. Task-completion latency is down more than 30% compared with the model it replaces in equivalent workloads, and agentic inference throughput is up to 2.5x faster per agent turn. For teams running multi-step agent loops, where a single user request can trigger a dozen or more model calls, that kind of per-turn speedup compounds fast. A workflow that took 40 seconds end to end could realistically drop into the teens.
Safety posture is the other notable shift. In malicious computer-use testing, Haiku 5.5 refused 82.59% of harmful tasks, up from 58.93% on Haiku 4.5. That score also beats Claude Sonnet 5.5 and Claude Opus 5.5, which both landed at 79.46% in the same test, though it still trails Claude Fable 5.1’s 87.50%. Anthropic’s system-card reporting places Haiku 5.5 at CB-1 and Autonomy-1 on its Responsible Scaling framework, meaning it doesn’t cross the CB-2 or Autonomy-2 thresholds that would trigger stricter deployment safeguards. The company is explicit that Haiku 5.5 remains broadly less capable than Claude Opus 5, which is by design: this is not a frontier reasoning model, it’s a volume workhorse.
One new wrinkle: Haiku 5.5 is the first Haiku model with built-in safeguards against a narrow set of high-risk cybersecurity requests. Anthropic says ordinary development tasks should be unaffected, but penetration-testing-style prompts and techniques more commonly associated with attackers face new restrictions. That’s a direct response to the kind of abuse patterns security researchers have flagged across the industry this year, including incidents covered in reports on AI models bypassing safety filters to write exploit code.
| Spec | Claude Haiku 4.5 | Claude Haiku 5.5 |
|---|---|---|
| Context window | 200,000 tokens | 1,000,000 tokens |
| Input price (under 100K tokens) | Higher baseline (per Anthropic) | $0.10 per million |
| Output price (under 100K tokens) | Higher baseline (per Anthropic) | $0.50 per million |
| Input price (over 100K tokens) | N/A | $0.50 per million |
| Output price (over 100K tokens) | N/A | $2.50 per million |
| Avg. cost to run vs predecessor | Baseline | ~75% cheaper |
| Malicious computer-use refusal rate | 58.93% | 82.59% |
| Agentic inference speed | Baseline | Up to 2.5x faster per turn |
| RSP classification | N/A | CB-1 / Autonomy-1 |
Why Anthropic is cutting small-model prices now
The timing isn’t an accident. Anthropic has been moving toward a public offering, and reporting on the company’s IPO materials has already flagged how central API revenue from exactly this kind of workload is to its business case. Our earlier coverage of Anthropic’s $2 trillion IPO filing and the follow-up piece on what the prospectus actually says both point to the same dynamic: Anthropic needs to show it can win and retain the high-volume, margin-thin segment of the API market, not just the premium reasoning tier where Opus competes.
That segment is also where Anthropic has been losing ground. A recent report detailed how Meta and Microsoft have been trimming their reliance on Claude models in favor of in-house alternatives, specifically citing cost as a factor for routine, high-volume tasks. A three-tier model lineup, Opus for frontier capability, Sonnet for general-purpose performance, Haiku for inexpensive high-throughput deployment, only works commercially if the cheap tier is genuinely cheap. Haiku 5.5’s pricing is Anthropic’s clearest signal yet that it understands this and is willing to compress margins on its smallest model to keep enterprise customers from defecting to cheaper in-house or open-weight alternatives.
There’s also a straightforward unit-economics argument. Small models cost less to serve per token, so even a 75% price cut doesn’t necessarily mean Anthropic is taking a 75% hit to gross margin on Haiku traffic, it likely means the model has gotten more efficient to serve, and Anthropic is passing some of that efficiency through to keep volume growing. Model providers across the industry have followed a similar pattern: ship a faster, cheaper small model roughly every two to three product cycles, because the small-model tier is where usage scales fastest and where competitive pressure on price is most intense.
How Haiku 5.5 pricing compares across the market
Available reporting does not include directly comparable benchmark scores pitting Haiku 5.5 against GPT-5 mini or Gemini Flash on standard coding and reasoning evaluations, so any claim about which small model “wins” outright would be speculation. What’s verifiable is the pricing structure itself, and how it sits relative to Anthropic’s own lineup. Haiku 5.5’s starting price of $0.10/$0.50 per million tokens is a steep discount against Sonnet 5.5’s $2/$10 per million tier, underscoring how much Anthropic wants Haiku positioned as the default choice for anything that doesn’t need Sonnet- or Opus-level reasoning.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Context window | Primary use case |
|---|---|---|---|---|
| Claude Haiku 5.5 (under 100K) | $0.10 | $0.50 | 1,000,000 tokens | High-volume, low-cost tasks |
| Claude Haiku 5.5 (over 100K) | $0.50 | $2.50 | 1,000,000 tokens | Long-document processing |
| Claude Sonnet 5.5 | $2.00 | $10.00 | Per Anthropic’s published tier | General-purpose reasoning |
| Claude Opus 5.5 | $4.00 | $20.00 | Per Anthropic’s published tier | Frontier-level reasoning |
For teams already running cost comparisons across the broader model market, our side-by-side on Sonnet 5.5 vs GPT-6 Luna vs Gemini 3.8 is a useful reference point for how Anthropic’s mid-tier pricing stacks up against OpenAI and Google. Haiku 5.5 effectively extends that competitive picture down into the budget tier, where the real battle over default-model selection in production apps is happening.
Market and enterprise impact
Because Anthropic is privately held, there’s no stock ticker to watch for an immediate market reaction. The more relevant signal is competitive: every major lab now has an incentive to respond with its own small-model price adjustment, a pattern that’s played out repeatedly since 2025 as OpenAI, Google, and Anthropic have traded blows on cost-per-token for their respective budget tiers. Enterprise procurement teams evaluating AI vendors increasingly treat small-model pricing as the deciding factor for high-volume contracts, since frontier-tier usage is typically a small fraction of total token spend compared to classification, extraction, and support automation.
That dynamic is already visible in how federal and state-level AI deployment has evolved. Projects like America.gov’s chatbot rollout across tens of thousands of sites depend on cheap-at-scale inference, not frontier reasoning, which is exactly the segment Haiku 5.5 is chasing. Regulatory attention adds another layer of pressure: the FTC’s ongoing probe into OpenAI and Anthropic AI agents means both companies are shipping new models under more scrutiny than at any point in the last two years, which likely explains why Anthropic is leading with safety metrics (the 82.59% refusal rate, the RSP classification) as prominently as it’s leading with price.
Historical context: the small-model price war
Haiku has always been Anthropic’s answer to the question of what happens when AI inference needs to run at consumer-app scale rather than research-lab scale. The original Haiku launched as a budget option; Haiku 3.5 and Haiku 4.5 each pushed the price-to-capability ratio further, and Haiku 5.5 continues that trajectory with the largest single jump in context window Anthropic has made at this tier. The pattern across the industry has been consistent: flagship models get more expensive and more capable at the top end, while budget models get cheaper and gain context length, because that’s where usage volume actually lives in production systems.
This also lines up with Anthropic’s broader 2026 product cadence. Three Claude 5.5 models in roughly a month, Opus, Sonnet, then Haiku, mirrors the staggered release pattern competitors have used, where a flagship model ships first to capture headlines and benchmark attention, and the budget tier follows once the underlying architecture improvements have been validated at the top of the lineup. It’s a reasonable bet that whatever efficiency gains let Opus 5.5 and Sonnet 5.5 ship were the same gains that made the 75% Haiku price cut possible without Anthropic taking on unsustainable serving costs.
What developers should actually do with this
For teams already on Haiku 4.5, the migration calculus is straightforward: same API shape, lower price, bigger context window, faster agentic turns. The main thing to test before flipping production traffic over is whether the new safety restrictions on cybersecurity-adjacent prompts affect any existing workload, particularly for teams building developer tools, code review assistants, or security tooling that might trip the narrower allowlist Anthropic has put in place.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-haiku-5-5",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "Summarize this support ticket."}]
}'Teams running classification or extraction pipelines that previously had to split documents to stay under Haiku 4.5’s 200,000-token ceiling can now test running the full document through Haiku 5.5’s 1 million-token window instead, which removes a layer of chunking logic and the associated error surface. That’s a real engineering simplification, not just a marketing number, since chunking logic is a common source of context-loss bugs in production RAG and extraction systems.
The three-tier strategy behind Opus, Sonnet, and Haiku
Anthropic’s model lineup now maps cleanly to three distinct buyer problems. Opus 5.5 targets teams that need the strongest reasoning available and will pay $4/$20 per million tokens for it. Sonnet 5.5 sits in the middle at $2/$10, aimed at general-purpose production use where quality still matters but cost discipline is part of the job. Haiku 5.5, at $0.10/$0.50 for shorter prompts, is built for the workloads that run millions of times a day and where a fraction-of-a-cent-per-call difference adds up to a real line item on a cloud bill. That segmentation isn’t unique to Anthropic, but the spread between Haiku and Opus pricing, a 40x difference on input tokens, is wider than it’s been at any point in Anthropic’s model history, which signals how aggressively the company is trying to own the low end of the market rather than cede it.
Open-weight models are the backdrop to this price cut
Haiku 5.5 isn’t launching into a vacuum. Open-weight releases have been closing the gap on small-model tasks over the past year, and that competition is arguably a bigger factor in Anthropic’s pricing decision than rivalry with OpenAI or Google alone. When a team can self-host a capable open model for the price of compute alone, a hosted API has to justify its premium with either meaningfully better quality, lower operational overhead, or a price that makes self-hosting not worth the engineering time. Haiku 5.5’s pricing reads like an attempt to make that calculation come out in Anthropic’s favor for as many teams as possible, especially smaller engineering shops that don’t want to run their own inference infrastructure.
Risks and open questions
Pricing announcements are easy to verify; real-world performance under production load takes longer to confirm. Anthropic’s 2.5x agentic speedup and 30% latency improvement are company-reported figures tied to specific internal workloads, not independently audited benchmarks, so teams migrating production traffic should run their own load tests before fully committing. The new cybersecurity-prompt restrictions are also untested at scale: it’s not yet clear from public reporting exactly where Anthropic has drawn the line between legitimate security tooling and restricted use, which is the kind of ambiguity that tends to generate false positives for security researchers and red teams doing legitimate work.
There’s also the question of how durable a 75% price cut is. Model providers have occasionally walked back aggressive introductory pricing once usage patterns settle, particularly if a model turns out to be used for far more compute-intensive tasks than the pricing tier assumed. Nothing in current reporting suggests Anthropic plans to revisit Haiku 5.5 pricing, but it’s worth flagging as the kind of thing that can change without much notice in a market this competitive.
Predictions: where this goes next
- Competitor price matching within weeks. OpenAI and Google have both responded to prior Anthropic small-model price cuts within a short window historically; expect a GPT-5 mini or Gemini Flash pricing adjustment to follow.
- Haiku becomes the default model in more agent frameworks. With a 1 million-token window at this price point, expect orchestration frameworks and agent SDKs to shift their default “cheap model” recommendation to Haiku 5.5 for routing and sub-task delegation.
- More scrutiny on the cybersecurity safeguards. Security researchers will likely probe the new restricted-prompt allowlist to understand exactly what’s blocked, similar to how researchers have tested other models’ safety boundaries this year.
- IPO narrative leans harder on volume metrics. Expect future Anthropic investor materials to cite Haiku-tier adoption and token volume growth as evidence of enterprise stickiness, separate from frontier-model headlines.
- Enterprise customers renegotiate contracts. Companies that moved workloads away from Claude toward in-house models, as Meta and Microsoft reportedly did, may find the new pricing enough to reopen that conversation, though switching costs for already-migrated infrastructure will slow any reversal.
The bigger picture for AI infrastructure spending
Haiku 5.5’s launch is a small line item next to the hundreds of billions being committed to AI data center buildouts, but it’s arguably more representative of where the money actually gets made. Frontier models generate headlines; small, cheap, fast models generate the token volume that pays the bills. A 75% price cut on the model built for exactly that volume signals Anthropic’s confidence that falling serving costs, not just competitive pressure, are driving this release. Whether that confidence is justified will show up in the next few quarters of usage data, not in this week’s announcement.
What’s clear today is that the gap between “frontier AI” and “cheap AI you can run everywhere” just got wider at the bottom end, in the customer’s favor. A model with GPT-4-class usefulness for narrow tasks, a million-token window, and a price under a dollar per million blended tokens changes the math for a lot of products that were previously too expensive to build on frontier APIs at all.
Frequently Asked Questions
When did Claude Haiku 5.5 launch?
Anthropic released Claude Haiku 5.5 on October 7, 2026, with immediate availability on Anthropic’s own platform plus AWS, Google Cloud, and Microsoft Azure.
How much does Claude Haiku 5.5 cost per million tokens?
For prompts under 100,000 tokens, it’s $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 tokens, pricing rises to $0.50 input and $2.50 output per million tokens.
How much cheaper is Haiku 5.5 than Haiku 4.5?
Anthropic says it costs approximately 75% less to run than Haiku 4.5 on average, which breaks down to a 90% cut for prompts up to 100,000 tokens and a 50% cut for prompts longer than that.
What is the context window for Claude Haiku 5.5?
Haiku 5.5 has a 1 million-token context window, up from 200,000 tokens on Haiku 4.5, a fivefold increase.
Is Claude Haiku 5.5 safer than previous Haiku models?
Anthropic reports it refused 82.59% of malicious computer-use tasks in testing, compared with 58.93% for Haiku 4.5. It also includes new safeguards for a narrow set of high-risk cybersecurity prompts.
How does Haiku 5.5 compare to Claude Sonnet 5.5 and Opus 5.5 on safety testing?
In the same malicious computer-use test, Sonnet 5.5 and Opus 5.5 both scored 79.46%, lower than Haiku 5.5’s 82.59%, while Claude Fable 5.1 scored highest at 87.50%.
Does Claude Haiku 5.5 cross Anthropic’s Responsible Scaling thresholds?
No. Anthropic classifies it at CB-1 and Autonomy-1, below the CB-2 and Autonomy-2 thresholds that would require stricter deployment safeguards.
Why is Anthropic cutting small-model prices now?
The cut comes as Anthropic positions itself ahead of a planned IPO and faces reported pressure from large customers shifting routine workloads to cheaper in-house alternatives, making the high-volume small-model tier increasingly important to its business case.
