Cost per token ranks list prices. Cost per task ranks what you actually pay to finish a job. After the Sep 2026 price war, cheaper $/1M stickers often lose when a model burns more output tokens, retries, or tool steps. Use TokenCalculator to hold one prompt fixed, then measure finished-task spend. Catalog rates checked 23 Sep 2026.
Price the finished job
Open the cost calculator, paste one golden prompt, set expected output tokens and monthly volume, then swap models. Do not declare a winner from the rate card alone.
Cost per token: what the rate card answers
Cost per token (usually shown as USD per 1M input and output tokens) answers: what does the provider charge for each meter line? It is the right tool for comparing stickers, Batch discounts, Fast mode multipliers, and cache read vs write lines.
| Model (standard band) | Input / 1M | Output / 1M | Cache read / 1M |
|---|---|---|---|
| GPT-6 Sol | $2.00/1M | $10.00/1M | $0.200/1M |
| Grok 4.7 (<200K) | $2.00/1M | $6.00/1M | $0.500/1M |
| Claude Opus 5.5 | $4.00/1M | $20.00/1M | $0.200/1M |
| GPT-6 Luna | $0.100/1M | $0.500/1M | $0.010/1M |
On stickers alone, Grok undercuts Sol on output ($6 vs $10) while matching Sol on input ($2). Opus 5.5 costs more on stickers than Sol. Luna sits far below all of them for volume chat. That ranking is not your monthly invoice.
Cost per task: what FinOps actually needs
Cost per task asks: what did we spend to reach an accepted answer, merge, or agent outcome? It includes input, output, reasoning or thinking tokens billed as output, cache writes on the first pass, retries, tool results stuffed back into context, and any long-context cliff that reprices the whole request.
Why cheaper tokens become bigger bills
- More output tokens per successful answer (verbose models, higher effort)
- More agent steps or tool loops before the task finishes
- Worse cache hit rate (fresh prefixes every turn)
- Hitting a long-context band (OpenAI 272K, xAI 200K) for the whole request
- Fast mode or priority tiers at about 2x standard
- Exact vs Approx token planning errors across vendors
Worked planning examples
Example A, equal tokens: 5K input + 2K output. Sol about $0.03. Grok about $0.022. Stickers favor Grok.
Example B, same task but Grok emits 4K output: Grok about $0.034, now above Sol $0.03. Cheaper token, bigger bill.
Example C, cache-heavy agent: 80K cached + 2K fresh + 1K output. Sol about $0.030. Grok about $0.050 because cache reads are $0.50 vs $0.20. The output sticker no longer decides the race.
How to measure cost per task in TokenCalculator
- Pick one golden prompt that represents production work
- Count tokens with Exact labels where available (OpenAI o200k); treat Claude and Grok as Approx until provider meters confirm
- Set output size to the median successful completion, not the marketing max
- Add cache hit share for agent prefixes
- Multiply by monthly task volume, not by raw API calls if one task spans many calls
- Re-run on Sol, Grok, Opus 5.5, and Luna before you lock a default
When stickers are enough
Use cost per token when workloads are short, uncached, and token counts stay similar across models (classification, light chat, fixed-size completions). Switch to cost per task for agents, coding loops, long RAG, and any path that can hit a context cliff.
Common mistakes
- Picking a vendor from a $/1M table without equalizing finished tasks
- Ignoring retries and tool transcript growth
- Comparing ChatGPT plan credits to API $/1M
- Budgeting Claude or Grok with OpenAI Exact counts
- Forgetting Batch vs Fast mode multipliers on the same sticker model
Frequently asked questions
What is cost per token vs cost per task?
Cost per token is the provider rate card (USD per 1M tokens). Cost per task is spend to finish an accepted outcome, including retries, tools, cache, and cliffs.
Why is a cheaper model more expensive on my invoice?
It likely burned more tokens per success, missed cache, hit a long-context band, or ran on a 2x Fast tier. Re-price a full successful trace.
Is Grok 4.7 cheaper than GPT-6 Sol?
On output stickers under 200K, yes. On cache-heavy or verbose tasks, Sol can win. See the Sol vs Grok guide.
Should I always pick the lowest $/1M model?
No. Pick the lowest cost per accepted task that clears quality gates. Luna can win volume chat while Sol or Opus wins hard agents.
How do I estimate agent cost before a run?
Count the stable prefix, expected fresh tokens per step, expected steps to success, and cache hit share. Price that trace in the cost calculator, then confirm with provider count APIs where Exact matters.
Do cache reads change the ranking?
Yes. Sol and Opus 5.5 both list $0.20 cache reads in this catalog while Grok lists $0.50 under 200K. Cache-heavy agents often reorder sticker winners.
Are TokenCalculator rates live?
No. Curated list prices last verified 23 Sep 2026. Confirm critical budgets on provider pricing pages.
What should I open first?
Paste one production prompt into the cost calculator on two candidate models, hold output fixed, then raise output to each model median success length and compare again.
Next steps
- GPT-6 Sol vs Grok 4.7
- GPT-6 Sol vs Claude Opus 5.5
- GPT-6 272K long-context pricing
- Prompt caching guide
- How prompt cost is calculated
- How we verify rates
Related Sep 2026 pricing guides
Frontier rate cards, cost-per-task math, and long-context cliffs. Cross-link these when you compare Sol, Luna, Opus 5.5, Grok 4.7, or Astra.