TokenPAPATokenPAPA
User GuideAPI ReferenceAI ApplicationsBlogPricingSign up

Claude 5.5 Pricing: Sonnet 5.5 vs Opus 5.5 vs Haiku 5.5

Sonnet 5.5 costs $1.86/$9.30 per 1M tokens and Opus 5.5 is exactly 2x that. Verified prices for all 11 Claude IDs on TokenPAPA, plus per-task cost math.

Claude 5.5 Pricing: Sonnet 5.5 vs Opus 5.5 vs Haiku 5.5

Claude Sonnet 5.5 costs $1.86 per 1M input tokens and $9.30 per 1M output tokens. Claude Opus 5.5 costs $3.72 and $18.60 — exactly twice Sonnet on both sides. Both are live on TokenPAPA, both accept the Anthropic Messages format, and both are cheaper than the models they replace.

That last part is the unusual bit. New flagship generations usually cost more than the ones they displace. This one went the other way: Sonnet 5.5 is 41.9% cheaper than Sonnet 4.6, and Opus 5.5 is 30.3% cheaper than Opus 4.6, on identical ratios in both directions. Nothing in the API changed to earn that discount — same endpoints, same request shape, same key.

This article is the pricing reference for that generation: every Claude ID on the platform, the per-task arithmetic, the break-even math for choosing between Sonnet and Opus, and the two request formats you can use.

Verified on 2026-11-11 against TokenPAPA's live pricing data (/api/pricing, 85 model entries) and the published model metadata. Prices are per 1M tokens in USD. Where a figure is an assumption rather than a published price, it is labelled as one.


Every Claude model on TokenPAPA, priced

Eleven Claude IDs are live. Sorted by input price:

Model IDInput / 1MOutput / 1MCached input / 1MEndpoints
claude-haiku-5-5$0.01$0.50$0.01anthropic, openai
claude-sonnet-5-5$1.86$9.30$0.10anthropic, openai
claude-sonnet-5$2.00$9.60$0.20anthropic, openai
claude-sonnet-4-6$3.20$16.00$0.32anthropic, openai
claude-opus-5-5$3.72$18.60$0.20anthropic, openai
claude-opus-5$4.65$23.25$0.50anthropic, openai
claude-opus-4-8$4.78$23.90$0.478anthropic, openai
claude-opus-4-7$4.78$23.90$0.478anthropic, openai
claude-opus-4-6$5.34$26.70$0.534anthropic, openai
claude-fable-5$9.562$47.812$0.956anthropic, openai
claude-fable-5-1$9.60$47.80$0.240anthropic, openai

Four patterns fall out of that table, and they matter more than the individual rows:

Output costs exactly 5× input on both 5.5 models. $9.30 ÷ $1.86 = 5.00. $18.60 ÷ $3.72 = 5.00. Any estimate you build should weight output tokens five times, not equally.

Opus 5.5 is exactly 2× Sonnet 5.5. Not "roughly". Both directions divide cleanly by two. There is no volume tier, no discount curve, no cleverness that changes that ratio — which means the Sonnet-versus-Opus decision is a pure question of how often Opus succeeds where Sonnet fails.

Sonnet 5.5 is priced identically to gpt-6.1-sol. Same $1.86 / $9.30, same $0.10 cached input. Two flagship-tier models from different vendors, one price point. If you were benchmarking OpenAI's new flagship against Anthropic's, the cost axis is now a tie and only quality differentiates.

Fable 5.1 and Fable 5 cost the same per token but not per cached token. Input differs by 0.4% ($9.60 vs $9.562), yet cached input is $0.240 versus $0.956 — four times cheaper on 5.1. If your workload has a stable system prompt or a fixed codebase prefix, that gap is worth more than the headline rate.

Key insight: Claude Sonnet 5.5 at $1.86 per 1M input and $9.30 per 1M output is 41.9% cheaper than Claude Sonnet 4.6 ($3.20 / $16.00), and its cached input rate of $0.10 per 1M is the same as OpenAI's gpt-6.1-sol. Opus 5.5 is exactly twice those numbers.

How much does a Claude 5.5 task actually cost?

Per-1M prices are hard to reason about. Here is the same data as per-call cost, using three request shapes that show up in real code:

  • Chat turn — 1,200 input tokens (system prompt + short history), 300 output tokens.
  • PR review — 12,000 input (diff + surrounding files), 1,500 output.
  • Agent step — 60,000 input (large working context), 2,000 output.
ModelChat turnPR reviewAgent step
claude-haiku-5-5$0.000162$0.000870$0.001600
claude-sonnet-5-5$0.00502$0.03627$0.13020
claude-opus-5-5$0.01004$0.07254$0.26040
claude-sonnet-4-6$0.00864$0.06240$0.22400
claude-opus-4-6$0.014418$0.104130$0.373800

At the PR-review shape and 1,000 calls per day:

ModelPer callPer dayPer month
claude-sonnet-5-5$0.03627$36.27~$1,088
claude-opus-5-5$0.07254$72.54~$2,176
claude-sonnet-4-6$0.06240$62.40~$1,872
claude-opus-4-6$0.10413$104.13~$3,124

The migration story writes itself. Moving that workload from claude-opus-4-6 to claude-opus-5-5 cuts the monthly bill from about $3,124 to about $2,176 — 30% off, same tier, no code change. Moving it from claude-sonnet-4-6 to claude-opus-5-5 costs more per call but buys a higher tier for 16% more than you were already paying for Sonnet.

And if you are currently on Opus 4.6 because you assumed Sonnet was not good enough, re-test. claude-sonnet-5-5 runs that same PR review for $1,088/month — 65% less than Opus 4.6 and 42% less than the Sonnet you might have rejected a generation ago.

Is Opus 5.5 worth exactly 2× Sonnet 5.5?

This is the only selection question that matters, and the 2× ratio makes it arithmetic rather than intuition.

Model the two strategies in units of "one Sonnet call" (S):

  • Always Opus: every request costs 2S.
  • Route: every request costs 1S. A fraction p of requests fail validation and escalate, and an escalated request costs S (the failed attempt) + 2S (the Opus retry) = 3S.

Setting the two equal: 1 + 3p = 2, so p = 1/3.

Escalation rateAlways OpusRoute (Sonnet first)Winner
10%2.00S1.30SRoute, 35% cheaper
20%2.00S1.60SRoute, 20% cheaper
33%2.00S2.00STie
50%2.00S2.50SOpus, 20% cheaper
80%2.00S3.40SOpus, 41% cheaper

Key insight: Because Opus 5.5 is exactly 2× Sonnet 5.5, routing to Sonnet first and escalating on failure is cheaper than always using Opus as long as fewer than one in three requests needs the escalation.

Two things that table hides, and you should not ignore them:

  1. Escalation has a latency cost, not just a token cost. A routed request that fails then retries pays two round trips. For interactive chat that is user-visible. For batch jobs it is invisible.
  2. Retry-on-failure requires a validator. You cannot route on "the model seemed unsure". You need a deterministic check — a schema validation, a test suite, a diff that applies cleanly, a numeric assertion. If you do not have one, you are measuring vibes, and the 33% threshold is not measurable either.

Where a validator does not exist, the cheaper move is usually to pick one model and stop optimizing. Pick Sonnet 5.5 for throughput, Opus 5.5 for correctness-critical work, and move on.

What changed between the 4.x generation and 5.5?

Both new models cut price on identical ratios in both directions, which means the discount is structural, not a promotional tweak to the input side.

ComparisonInput changeOutput changeRatio
claude-sonnet-4-6 → claude-sonnet-5-5$3.20 → $1.86$16.00 → $9.300.581 (41.9% cut)
claude-opus-4-6 → claude-opus-5-5$5.34 → $3.72$26.70 → $18.600.697 (30.3% cut)
claude-sonnet-5 → claude-sonnet-5-5$2.00 → $1.86$9.60 → $9.300.930 (7.0% cut)
claude-opus-5 → claude-opus-5-5$4.65 → $3.72$23.25 → $18.600.800 (20.0% cut)

The last two rows are the useful ones for anyone already on the 5 generation. Upgrading from claude-sonnet-5 to claude-sonnet-5-5 is a 7% discount; upgrading from claude-opus-5 to claude-opus-5-5 is a 20% discount. Neither requires a request change — same endpoint, same parameters, same key, a different model string.

Per TokenPAPA's model metadata, the 5.5 models are positioned for complex agentic coding, long-horizon reasoning and codebase-scale migrations, and list a 1M-token context window. That context ceiling is the reason the agent-step scenario above uses 60,000 input tokens rather than a small prompt: at this tier, input volume — not output — is what drives the invoice.

Do the 5.5 models speak the Anthropic Messages API?

Yes, and this is worth checking explicitly because it is the difference between a drop-in migration and a rewrite.

Every Claude ID on the platform lists both anthropic and openai endpoint types. That means two request formats against the same key and the same balance:

Anthropic Messages format — POST /v1/messages, requires an anthropic-version header:

TOKENPAPA_API_KEY="sk-..."

curl https://api.tokenpapa.ai/v1/messages \
  --oauth2-bearer "$TOKENPAPA_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5-5",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Summarise this migration plan in 5 bullets: ..."}
    ]
  }'

OpenAI-compatible format — POST /v1/chat/completions, no extra headers:

from openai import OpenAI

client = OpenAI(
    api_key="sk-...",
    base_url="https://api.tokenpapa.ai/v1",
)

resp = client.chat.completions.create(
    model="claude-sonnet-5-5",       # or "claude-opus-5-5"
    messages=[{"role": "user", "content": "Review this diff for race conditions: ..."}],
    max_tokens=1500,
)

print(resp.choices[0].message.content)
print(resp.usage)                    # input/output token counts for cost tracking

Both routes hit the same models at the same prices. The practical rule: keep whatever format your SDK already speaks. If you are on the Anthropic SDK, use /v1/messages; if you are on the OpenAI SDK, use /v1/chat/completions. Switching formats to save money does not work — there is no price difference between them.

Key insight: All 11 Claude model IDs on TokenPAPA support the native Anthropic Messages endpoint and the OpenAI-compatible endpoint under one key, so migrating from a direct Anthropic account does not require changing your client library — only your base URL and model string.

What does prompt caching change?

Cached input is where Claude pricing gets materially cheaper, and the 5.5 generation kept the discount while cutting the base rate.

ModelCached input / 1MCache ratio vs input
claude-sonnet-5-5$0.10~5.4% of input price
claude-opus-5-5$0.20~5.4% of input price
claude-sonnet-4-6$0.3210% of input price
claude-opus-4-6$0.53410% of input price

Rerunning the PR-review workload with 80% of input served from cache:

ModelNo cache80% cachedSaving
claude-sonnet-5-5$1,088/mo~$581/mo46.6%
claude-opus-5-5$2,176/mo~$1,162/mo46.6%
claude-sonnet-4-6$1,872/mo~$1,043/mo44.3%
claude-opus-4-6$3,124/mo~$1,740/mo44.3%

Two caveats that decide whether you actually see these numbers:

  • Caching only pays on a stable prefix. A system prompt in the same position, a fixed document corpus, an unchanged repository preamble. If your prompt is rebuilt differently each call, the cache never hits and the 46.6% is fiction.
  • Maximise the cached share, not the model choice. At 80% cache hit rate, claude-sonnet-4-6 ($1,043/mo) lands close to uncached claude-sonnet-5-5 ($1,088/mo). Two levers that look like one decision.

Which Claude model should you route each workload to?

WorkloadModelWhy
Classification, tagging, routingclaude-haiku-5-5Cheapest listed tier; short outputs
Chat and support repliesclaude-sonnet-5-55× output ratio keeps short answers cheap
RAG answer synthesisclaude-sonnet-5-5Large input, small output — input-heavy shapes favour it
PR review, code explanationclaude-sonnet-5-5$0.036 per review at the 12k/1.5k shape
Codebase-scale migration planningclaude-opus-5-5Long-horizon planning where a wrong plan is expensive
Multi-hour autonomous agent runsclaude-opus-5-5 or claude-fable-5-1Escalating cost is trivial next to a failed run
Workloads with huge cached prefixesclaude-fable-5-1$0.240 cached input vs $0.956 on claude-fable-5
Anything already working on 4.6claude-sonnet-5-5Same request shape, 41.9% lower rate

Haiku 5.5: $0.01 / $0.50, and how to use it safely

claude-haiku-5-5 is live at $0.01 input / $0.50 output per 1M tokens. The listing has been verified against the live pricing endpoint and re-confirmed on 2026-11-11 — the number is not a placeholder. That is ~186× cheaper on input than Sonnet 5.5 and ~18.6× cheaper on output, and its cached-input rate equals its input rate, so there is no separate caching discount.

Two things follow from that ratio. First, the 50:1 output-to-input ratio behaves very differently from the 5:1 ratio every other Claude ID on the platform uses: budget on output tokens, because an input-only estimate will badly understate the bill. Second, since caching buys nothing extra here, prompt caching is a Sonnet / Opus optimisation, not a Haiku one. Route high-volume, short-output work — classification, extraction, tagging, routing — to Haiku, and keep long-form generation on Sonnet 5.5.

What this pricing does not tell you

Being explicit, because "which is cheaper" is the wrong question on its own:

  • Token price is not cost per solved task. A model that needs two attempts is not cheaper at half the rate. Measure cost per successful outcome, not per call.
  • Quality differences between Sonnet 5.5 and Opus 5.5 are not in this article. I have verified prices, not benchmark scores. Do not read the 2× ratio as a 2× capability claim.
  • Reasoning tokens are billed as output. If these models emit internal reasoning, it lands in your output count at the output rate — five times the input rate. Budget on usage.total_tokens, not on the visible answer length.
  • Cached pricing assumes a cache hit. Cache writes and misses are billed differently. The 46.6% figure is an 80%-hit-rate scenario, not a guarantee.
  • Hosted API, not self-hosting. Standard data-handling implications apply for regulated workloads; see the privacy policy.

FAQ

How much does Claude Sonnet 5.5 cost?

$1.86 per 1M input tokens, $9.30 per 1M output tokens, $0.10 per 1M cached input tokens. That is 41.9% cheaper than claude-sonnet-4-6 on both input and output, and identical to the price of gpt-6.1-sol.

How much does Claude Opus 5.5 cost?

$3.72 per 1M input, $18.60 per 1M output, $0.20 per 1M cached input. Exactly 2× claude-sonnet-5-5 in both directions, and 30.3% cheaper than claude-opus-4-6.

Is Opus 5.5 worth twice the price of Sonnet 5.5?

Only if more than one third of your requests genuinely need escalation. Routing normally to Sonnet 5.5 costs 1 unit and escalating costs 3 units (the failed attempt plus the Opus call), against 2 units for always using Opus. Routing wins below a 33% escalation rate — and only if you have a deterministic validator to detect failure.

Do Claude 5.5 models work with the Anthropic Messages API format?

Yes. Every Claude ID on TokenPAPA lists both anthropic and openai endpoint types. Use POST /v1/messages with an anthropic-version header, or the OpenAI-compatible chat completions route — same key, same prices, no format penalty.

What is Claude Haiku 5.5 priced at?

$0.01 per 1M input and $0.50 per 1M output as listed on 2026-11-11 — a 50:1 output ratio with no caching discount. Verify the current rate on the pricing page before sizing a workload on it.

Do I need a new API key for the Claude 5.5 models?

No. All Claude IDs are enabled on default, Starter, Pro and Enterprise, so your existing key, base URL and balance work as-is. Change the model string and nothing else.

Where can I compare Claude against other vendors at the same time?

The pricing page carries the full live table across all 85 model entries on the platform, and the pricing cheat sheet groups them by tier. For cross-vendor comparisons, see GPT-6.1 Sol vs GPT-6 Astra and DeepSeek V4 vs GPT-5.6.


Start

Both 5.5 models are live now and share a key with the rest of the catalogue. Live pricing: https://tokenpapa.ai/pricing. Get a key at https://tokenpapa.ai/keys, or create an account at https://tokenpapa.ai/register and point your client at https://api.tokenpapa.ai/v1 — then change one string:

model="claude-sonnet-5-5"   # was: claude-sonnet-4-6

If you are sizing a coding workload rather than a chat workload, the code review cost breakdown works through the same arithmetic across DeepSeek, Qwen, Kimi and Claude.

Last updated 2026-11-11. All prices verified against TokenPAPA's live pricing data (85 model entries) on that date; ratios and monthly totals are arithmetic from those published rates. Per-task token counts are stated assumptions, not measured benchmarks. Quality claims are deliberately absent — this is a pricing reference, not an evaluation.

How is this guide?

Last updated on

Claude 5.5 Pricing: Sonnet 5.5 vs Opus 5.5 vs Haiku 5.5 | TokenPAPA