GPT-Image 2.6 Flare & Sunburst are liveImages from $0.025 each

Claude Haiku 5.5 API

NewLLM

The Claude Haiku 5.5 API is the fast Claude route with 1M context, vision, tools and adaptive thinking, launched at 20% below Anthropic's official token rates.

  • Input · <=100K$0.08 USDOfficial price$0.1 USD
  • Output · <=100K$0.4 USDOfficial price$0.5 USD
  • Input · >100K$0.4 USDOfficial price$0.5 USD
  • Output · >100K$2 USDOfficial price$2.5 USD
  • Savings20%
  • Context1M
  • View tool pricing and billing rules

What are we making today?

Bring an idea, a question, or a first draft.

Enter to send · Shift + Enter for a new line

01

What this model is for

The Claude Haiku 5.5 API is the fast Claude route with 1M context, vision, tools and adaptive thinking, launched at 20% below Anthropic's official token rates.

Claude Haiku 5.5 is the small, quick Claude that Anthropic released on October 7, 2026. The catalog gives it a 1M-token context window, text and image input, tool use, adaptive thinking, and a fast-tier mark, all at the lowest token rates in the Claude line. The jump from Haiku 4.5 is mostly capacity: the older model stops at 200K context, and this one also adds an effort dial that the Haiku family did not have before.

Put Haiku 5.5 where the job is narrow and the check is cheap: labelling tickets, pulling fields out of documents, compacting a long session so a larger model can continue, or acting as a subagent while Opus 5.5 or Sonnet 5.5 holds the plan. Two habits keep the bill predictable. Keep the prompt below the 100K input band, and recount tokens, because the tokenizer now spends more of them on the same text than Haiku 4.5 did.

Choose it for high-volume, latency-sensitive agent calls that still need long context, then compare it with Haiku 4.5 before switching.

Best-fit coding workloads

Ticket triage into fixed labels

Classify incoming support or bug tickets against a closed label list and return the label plus the sentence that decided it. A bounded output with a fixed vocabulary is easy to score, so a rubric plus a sample of hand-labelled tickets gives you an accuracy figure before the volume arrives.

Invoice fields into JSON

Extract vendor, date, currency, line totals and tax from scanned invoices and return a fixed schema, with null where a field is absent. Use strict tool definitions so the shape is enforced, and keep each page in its own request so the prompt stays inside the cheaper input band.

Context compaction for a longer run

Summarise a growing agent session into the facts the next step needs: decisions taken, files touched, open failures and the exact commands that were run. Ask for a fixed section order and a stated token target, then check that the following step still has every fact it requires.

02

Hands-on notes

Checked on 2026-10-27 through the OmniaKey API with one standard coding prompt (median of 3 runs) and three capability probes.

Output tokens for our standard coding prompt, reasoning included
501
Cost of that output at current prices
$0.0002
JSON mode
Passed
Tool calling
Passed
Image input
Passed

03

Technical specification

Catalog facts and integration details verified for this route.

Model ID
claude-haiku-5-5
Context
1M
Inputs and tools
Vision · Tool use · Fast responses · Adaptive thinking
API endpoint
/v1/chat/completions
Cache hit / 1M
$0.008 USD
Coding clients
Claude Code · Cursor · Cline · Aider

Limits to plan around

  • Input length selects one of two rate bands. A request that passes 100K input tokens is charged at the higher band for the whole request, and cache reads and writes count toward that length, so a long cached prefix does not stay in the cheap band.
  • The tokenizer spends roughly a third more tokens on the same text than Haiku 4.5, so budgets, log parsers and usage alerts sized on the older model will understate consumption after the switch.
  • A final assistant turn used as a prefill is rejected, and non-default temperature, top_p or top_k values are too. Constrained output has to come from tools or structured outputs instead of a primed reply.
  • A refusal can arrive as a normal response with stop_reason refusal and no fallback from the server, so a client that only inspects the HTTP status may log an empty answer as success.

04

Compare nearby models

Compare workload fit and current listed token rates before choosing a route.

ModelBest forContextInputOutput
Claude Haiku 5.5This page
claude-haiku-5-5
Narrow, repeatable jobs at volume where a rubric or schema can score every answer and prompts stay short.1M
$0.08 USD
Official price $0.1 USD
$0.4 USD
Official price $0.5 USD
Claude Haiku 4.5
claude-haiku-4-5
Existing 200K-context integrations where rewriting the request shape is riskier than keeping the older model.200K
$0.28 USD
Official price $1 USD
$1.4 USD
Official price $5 USD
Claude Sonnet 5.5
claude-sonnet-5-5
Sustained planning, multi-file changes and reviews where a mistake would cost more than the tokens saved.1M
$0.56 USD
Official price $2 USD
$2.8 USD
Official price $10 USD
GPT-6 Luna
gpt-6-luna
Long prompts above the Haiku input band, or stacks already built on OpenAI's Responses tooling.1M
$0.2 USD
Official price $0.1 USD
$1.2 USD
Official price $0.5 USD

05

SDK quickstarts

Use one base URL and model ID from cURL, Python, or TypeScript.

Streaming chat completion
curl https://api.omniakey.com/v1/chat/completions \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-haiku-5-5",
    "messages": [
      {"role": "user", "content": "Review this pull request"}
    ],
    "stream": true
  }'

06

Estimate token cost

Enter expected input and output tokens to compare current listed rates.

OmniaKey estimate
$0.016
Official-rate estimate
$0.02
Estimated savings
$0.004

Estimate covers input and output tokens only. Cache usage, retries, tools, and client-side token accounting can change the final charge.

08

Sources and verification

Model facts and links were checked on 2026-10-27. Provider documentation remains the authority for upstream behavior.

Frequently asked questions

How much does Claude Haiku 5.5 cost on OmniaKey?

Rates vary with input context length: <=100K: Input $0.08 USD, Output $0.4 USD, Cache hit / 1M $0.008 USD; >100K: Input $0.4 USD, Output $2 USD, Cache hit / 1M $0.04 USD.

What is the context window for Claude Haiku 5.5?

The listed context window is 1M. Actual usable limits can also depend on the client and upstream request format.

When was Claude Haiku 5.5 released?

Anthropic released Claude Haiku 5.5 on October 7, 2026, with claude-haiku-5-5 as its Claude API model id. The dated alias used by the previous generation does not exist for this release, so pin the id exactly as shown rather than extending the older pattern.

How does the 100K input band work?

Each request falls into one of two bands by input length: up to 100K tokens, then above it. Passing 100K moves the entire request to the higher input, output and cache-read rates, not just the excess tokens. Cache reads and writes are counted when the prompt length is measured, so a cached prefix can still push a request over the line.

Is Claude Haiku 5.5 a good fit for latency-sensitive agents?

The catalog marks it as a fast tier, and it is the smallest model Anthropic offers in the 5.5 generation. The lever you control is output_config.effort: a medium setting with a tight max_tokens keeps each step small. Measured completion time on your own calls is the only figure worth planning around, and it depends on tools and prompt length.

How do I call Claude Haiku 5.5 through the API?

Send POST /v1/chat/completions to https://api.omniakey.com/v1 with claude-haiku-5-5 as the model and any OmniaKey key. Anthropic's native shape also accepts output_config.effort and a thinking object set to adaptive. In Claude Code, point ANTHROPIC_BASE_URL at https://api.omniakey.com and set ANTHROPIC_AUTH_TOKEN to the key.

Is Claude Haiku 5.5 good for coding?

It suits the small steps around a coding session: locating a symbol, triaging a failing test, making a mechanical edit that a suite protects, or summarising a long log. Anthropic still points complex, open-ended agentic coding at Sonnet 5.5 and Opus 5.5. Keep a deterministic check on every change before you let the cheap tier edit files unattended.

Can I reuse my Haiku 4.5 request settings?

Not unchanged. Replace manual enabled thinking and budget_tokens with adaptive thinking plus effort, remove assistant prefills and old sampling values, parse blocks by type instead of position, and swap the old computer-use tool for the current one. Also recount your token budget, because the same text now consumes more tokens.