TokenAtlasClaude pricing in one line
Sonnet 3.5: $3 in / $15 out per 1M tokens. Opus: $15 in / $75 out. Haiku: $0.25 in / $1.25 out. Output tokens dominate cost — typically 60–80% of the bill on long-form generation.
Worked example
A long-context summarization job processing 200 documents/day at 8K input + 1K output tokens runs ~48M in / 6M out per month. On Sonnet that's $144 + $90 = $234/mo. On Haiku, $12 + $7.50 = $19.50/mo.
When Claude beats GPT
On 100K+ context windows, large-document QA, and structured extraction, Sonnet 3.5 often matches GPT-4o quality at competitive cost. Use the calculator side-by-side with the OpenAI calculator to confirm.
Cost drivers ranked for Claude workloads
Output tokens dominate a Claude estimate because Sonnet bills output at 5x input ($15 vs $3 per 1M) and Haiku at 5x as well ($4 vs $0.80). After output length, the next driver is context size carried on every turn — long documents re-sent each call multiply input cost even when the question is short. Call volume is the third lever, and it is the only one that scales the estimate without changing per-call economics.
Sonnet vs Haiku for the same workload
Blended at 70/30: Sonnet = 0.7x$3 + 0.3x$15 = $6.60/1M tokens. Haiku = 0.7x$0.80 + 0.3x$4 = $1.76/1M. On a 10M input / 4M output monthly workload, Sonnet models to $30 + $60 = $90; Haiku to $8 + $16 = $24. Run both scenarios side by side before assuming the higher-quality model is required for the task.
Assumptions to write down before comparing models
A Claude estimate is only comparable to another provider's if the input/output split, context length and monthly call count are stated the same way for both. Note whether the workload is single-turn or multi-turn (context re-billing changes input volume sharply) and whether output is capped, since these three assumptions explain most of the gap between two runs of the calculator.
A second scenario: high-volume support triage
A support team routing 300,000 tickets/month through Claude for triage at 600 input / 120 output tokens each is 180M input and 36M output tokens. On Haiku (0.7x$0.80+0.3x$4=$1.76/1M blended): ~216M total x $1.76/1M ≈ $380/month. On Sonnet at $6.60/1M blended, the same volume models to roughly $1,426/month — a gap worth checking before defaulting to the higher-quality model for a routing task.
Re-running the estimate when Anthropic changes rates
Anthropic has adjusted per-token rates and added tiers before, so treat any Sonnet/Opus/Haiku figure here as tied to the rate table at the top of this page. When a rate changes, re-enter the same input/output/volume numbers rather than re-deriving the workload from scratch — the token counts don't change, only the multiplier, so the fastest check is a straight substitution into the same formula.
What this calculator does not do
It prices the token volumes you enter against Anthropic's published per-token rates; it does not read your Anthropic console, reconcile an invoice, or apply org-specific negotiated discounts. Bedrock and Vertex multipliers are separate inputs, not automatic lookups. Treat the output as a planning estimate to compare models and scenarios, not as a substitute for the billing page.
Frequently asked questions
- Is prompt caching included?
- Yes — Claude's 90% cache discount is modelled when caching is enabled.
- Does it cover Bedrock and Vertex pricing?
- Direct Anthropic API rates are default; Bedrock and Vertex multipliers are on the Pro plan.
- How are batch requests priced?
- Anthropic's 50% Batch API discount applies when you toggle 'batch mode'.