Price per Request, or Cost per Completed Task

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. Claude Haiku 4.5 costs $1 and $5, so on the price list the gap is tenfold. It is tempting to read that as a bill cut to a tenth.

It does not work that way. In Anthropic’s published measurements, Haiku 4.5 answered knowledge questions at about a tenth of Claude Opus 5’s cost per question, at 63% accuracy. Opus 5 scored 92% on the same set. A failed request still bills its tokens, then the retry bills again, and then whatever the failure costs downstream bills as well. So the unit of comparison is not price per request but cost per completed task.

Addy Osmani put the confusion precisely in July 2026. “The tier structure itself is clear, but the matrix beneath it, model by reasoning effort, is not.” What practitioners are stuck on is not the order of magnitude of the price but that two-dimensional table.

What Moved in the Past Year

Date	Event
2025-10-01	Claude Haiku 4.5 released ($1/$5, 200K context)
2026-06-09	Claude Fable 5 and Claude Mythos 5 announced. The first widely released models with thinking always on
2026-06-12	US export controls suspend Fable 5 and Mythos 5 for all users
2026-06-26	OpenAI previews GPT-5.6 Sol under government access restrictions
2026-06-30	Controls lifted; Fable 5 redeployed on 2026-07-01
2026-07-09	GPT-5.6 Sol, Terra and Luna released broadly
2026-07-24	Claude Opus 5 announced ($5/$25, the same price as Opus 4.8)
2026-08-05	Claude Opus 4.1 retired, replaced by Opus 4.8
2026-08-21	GPT-5.6 Sol price cut from $5/$30 to $4/$20
2026-10-23 (planned)	OpenAI retires gpt-4-turbo, gpt-4o, o1, o3-mini and others, consolidating onto Sol, Terra and Luna

What Is on the Table

Dollars per million tokens, retrieved 2026-08-29.

Model	Input	Output	Context	Max output	Vendor’s own positioning
Claude Fable 5	10.00	50.00	1M	128K	The most demanding reasoning and long-horizon agentic work
Claude Opus 5	5.00	25.00	1M	128K	Multihour autonomous coding agents, large-scale refactoring
Claude Sonnet 5	2.00	10.00	1M	128K	Frontier intelligence at scale; daily coding
Claude Haiku 4.5	1.00	5.00	200K	—	Near-frontier performance at the most economical price point
GPT-5.6 Sol	4.00	20.00	1.05M	128K	Flagship model for complex professional work
GPT-5.6 Terra	2.00	12.00	—	—	Balances intelligence and cost
GPT-5.6 Luna	0.20	1.20	—	—	Optimized for cost-sensitive workloads
Gemini 3.7 Flash	0.75	3.75	—	—	Introductory pricing through 2026-12-31
Gemini 3.5 Flash-Lite	0.30	2.50	—	—	A workhorse built for cost efficiency and high volume

Sonnet 5 was scheduled to rise to $3/$15 on 2026-09-01, but the pricing page now states that the increase will not happen and $2/$10 is the standard rate. Sol’s $4/$20 is explicitly a promotional price available at least through 2026-11-21. Sol also carries a long-prompt surcharge: prompts above 272,000 input tokens bill input at 2x and output at 1.5x. Gemini 3.1 Pro Preview has a similar two-step surcharge at 200,000 tokens. The Claude models currently have no such surcharge.

Caching and batching rates are nearly identical across the two vendors. Prompt-cache reads cost 0.1x standard input on both, and writes cost 1.25x. Both Batch APIs are 50% off, and on the Anthropic side that discount stacks on top of cache reads and writes.

The Order the Vendors Actually Published

Anthropic’s model-selection guide says to sweep effort before touching the model. “Tuning effort is often a better lever than switching models.” The cost-optimization guide goes further and fixes the order by splitting the levers in two.

• Free wins: prompt caching, input-token hygiene, loop hygiene, output hygiene, batch processing. They lower the bill without lowering output quality, and caching stays on permanently.
• Tradeoffs: task budgets, effort, model selection, multi-model architectures. Each exchanges capability for cost, so they come after the free wins are exhausted.

OpenAI lands in the same shape. reasoning.effort has six steps, none / low / medium / high / xhigh / max, with medium as the default. The migration instruction is to start from your current GPT-5.5 or 5.4 setting, then test that same setting and one level lower. New projects start at medium, and latency-sensitive workloads at low. Output length is controlled separately through verbosity (low / medium / high), and difficult tasks can additionally set reasoning.mode to pro.

Anthropic’s published numbers show that this order is a measurement rather than a maxim.

Workload	Setting	Accuracy	Cost
Research work (Fable 5)	low	1–3 points below default	a third to a half
Research work (Fable 5)	medium	matches default	70–85%
Long-horizon coding (Opus 5)	medium	about 2 points below	half
Long-horizon coding (Opus 5)	low	about 8 points below	a quarter
Deep multi-subtopic research	each step	about 2.4 rubric points per step	no free cut

On research work, the default effort bought nothing measurable over medium on any of the four benchmarks. Time per problem also fell from 7.9 minutes to 4.5. Long-horizon coding, by contrast, is a real tradeoff. Which curve applies is a property of the workload, so someone else’s numbers cannot be transplanted into your configuration.

Where failures are detectable, there is a cheaper shape still. Running everything at low and re-running only the failures at the default effort solved about 93% of tasks at roughly $0.70 each. Running everything at the default solved 91.7% at $1.39. Counting the wasted cheap attempts, the same or better pass rate came in at half the price. Starting at medium solved about 94% at roughly $0.95. This shape needs a usable failure signal (tests, a checker, a validator), and the cost of that checker plus the doubled wall-clock on failures belong in the arithmetic.

Where Fable 5 Fits

The first thing that matters about Fable 5 is not its price but that the API surface changed.

• Thinking is always on: the correct move is to omit the thinking parameter. Both {"type":"disabled"} and budget_tokens return a 400. Depth is controlled only through output_config.effort.
• The raw chain of thought is never returned: display: "summarized" gives a readable summary, but the default is "omitted" and the thinking text arrives empty. Set it explicitly if you render reasoning to users.
• 30-day data retention is required: requests from a zero-data-retention organization fail with a 400.
• Turns run long: a single request on a hard task running fifteen minutes is described as normal. Plan timeouts, streaming and progress indicators before migrating.

The use the vendor names is work that does not fit in one sitting. The Claude Code documentation calls out root-cause investigations, outage debugging and architecture decisions, and advises handing it the result you want and letting it plan the path.

Low effort deserves particular care here. Anthropic states that lower effort settings, including low, still perform very well on Fable 5, often exceeding the xhigh or even max performance of previous models. That puts “Fable 5 at low” in direct competition with “an older model at high effort.” It is the reason to measure the expensive model at reduced effort before dropping to a cheaper one. Caches are model-scoped, so mixing models also forfeits cache reuse, which pushes in the same direction.

Prompting changes too. Prompts and skills written as step-by-step scaffolding for earlier models reduce Fable 5’s output quality. The official recommendation is to state goals and constraints instead, then A/B with the old scaffolding removed. At higher effort it will sometimes start tidying or abstracting things nobody asked for, so boundaries have to be stated. Parallel sub-agents are dependable, and sub-agents that communicate asynchronously outperform spawn-and-block: long-lived agents keep their context, so cache reads land, and the orchestrator is not bottlenecked on the slowest one.

Where GPT-5.6 Sol Fits

Sol’s distinguishing specifications are an asymmetric window, 1,050,000 tokens of context of which only 922,000 can be input, and a price step at 272,000 tokens. Its knowledge cutoff is 2026-02-16. Under the Preparedness Framework, all three of Sol, Terra and Luna are rated “High” for Cybersecurity and for Biological/Chemical work, reportedly the first time smaller and faster models in the same family received that rating.

Two lines in the System Card carry operational weight. The first is that Sol shows a greater tendency than GPT-5.5 to go beyond the user’s intent, with unauthorized actions occurring more often (the absolute rate is noted as low). The second is that using the internal coding agent across long sequences of tasks requires active user supervision. That sits badly with a design that runs unattended.

On controls, Sol is finer-grained than Fable 5. Effort has six steps, including a none step that skips thinking entirely. Output length is set independently through verbosity, and since GPT-5.6 is more concise by default than 5.5, migrations sometimes need verbosity raised rather than lowered. Ultra mode goes beyond a single-agent setup, running four sub-agents in parallel by default, trading token consumption for quality and speed on complex work, which points the same way as Fable 5’s asynchronous delegation.

Below Sol sit Terra and Luna in the same family. Luna’s $0.20 input is a twentieth of Sol’s, and staying inside one family makes the tier easier to move without rewriting prompts.

When Two Models Beat One

Anthropic names exactly two multi-model shapes.

• Advisor: a cheaper executor runs the loop and consults a frontier model on hard decisions.
• Orchestrator: a frontier model plans and delegates bulk work to cheaper workers.

Both pay only under narrow conditions.

The advisor pays when the capability gap is wide and the executor actually consults. The consult rate is the fragile variable: lowering effort can drop a pairing from consulting on most tasks to almost none, at which point it scores below the executor alone. Gating the consult well requires a cheap signal, because asking the executor to recognize the hard cases itself demands the very judgment it is missing. On the coding benchmark, the flagship pairing was the most accurate configuration measured but sat within noise of the frontier model alone at medium effort, at about the same cost.

The orchestrator pays when there is bulk to hand off. On work larger than any context window it cost 55% less than the frontier model solo at every effort setting, scoring 3 to 7 points below its best. On routine search work it paid as tail insurance, about half the average cost and a third at the 90th percentile, but it reversed on the harder full set. When the work is one dependent chain, or fits in a single context window, the orchestrator pays for a plan, a handoff and a merge that a single model gets for free. In every such case measured, the coordinator’s own model alone at lower effort came out ahead.

All of this sits on top of caching and batching. In an agent loop, cache hit rates of 81% to 90% produce a factor of 2.5 to 3.7 in cost. The Batch API is 50% off and stacks on the cache. Checking whether those are in place comes before moving any model.

The Ground the Argument Stands On

The gaps between models are themselves moving.

The UK AI Security Institute, having evaluated more than thirty frontier systems continuously over two years, reports that the capability gap between open-weight and closed frontier models has narrowed to four to eight months. The length of task a model can complete autonomously grew from under five minutes at the end of 2023 to over sixty minutes by mid-2025, doubling roughly every eight months. The inference cost of reaching GPT-3.5-level performance fell by a factor of 280 between November 2022 and October 2024.

Meanwhile the instruments of comparison are themselves under suspicion. SWE-bench Verified solve rates rose from about 40% at the end of 2024 to over 60%, but verbatim overlap with benchmark data was found in coding problems, and performance collapsed when the problems were paraphrased. Osmani’s warning about models gaming their own evals rests on the same worry. If public benchmarks cannot stand in for your workload, the switching decision comes back to an evaluation you build yourself.

Falling unit costs also do not automatically mean falling bills. Gartner projects that inference cost for trillion-parameter models will drop more than 90% by 2030 against 2025, while noting that agentic usage consumes 5 to 30 times the tokens per query [primary source unverified]. McKinsey attributes 60% of agentic AI cost to response refinement and argues that a higher-priced frontier model is often the rational choice because it shortens verification time by more than the added cost [primary source unverified].

What Enterprises Actually Do

Running several models is now the norm rather than the exception. In a16z’s CIO survey, the share of companies using five or more models rose from 29% to 37%, and the stated reason was per-use-case performance differences rather than avoiding vendor lock-in. In LangChain’s survey of 1,340 practitioners, more than three in four run multiple models in development and production, and quality (32%) ranks above cost among priorities, with cost mentioned less often than the year before.

Adoption of the cheaper tiers splits sharply by vendor. Non-frontier models reach production at 67% for OpenAI and 41% for Google, but only 27% for Anthropic. In OpenRouter’s data across roughly 100 trillion tokens, the median effective cost is $0.73 per million tokens, and the correlation between price and usage is weak enough that demand reads as price-inelastic. Making a cheaper model available does not, by itself, move traffic to it.

In the same data, reasoning models grew from a negligible share of tokens in early 2025 to over 50% by late in the year, and programming’s share of tokens grew from 11% to over 50% across the year. The whole distribution is moving toward using expensive thinking models for code.

Where Practitioners Draw the Line

Where the official guides put effort first, practitioners describe a division of roles.

Simon Willison assigns Fable 5 to judgment and audit, Sonnet to implementation, and Haiku to mechanical edits. His reasoning is plain: implementation work rarely needs a top-tier model. Adding a single sentence to Claude Code, telling it to decide on an appropriate lower-powered model for every coding task and run it in a sub-agent, slowed his Fable burn rate without reducing how much work got done. Only judgment, review and integration stay in the main loop.

The same writer, on Sol, reports not finding it better than Fable for the kind of complex coding he does with Fable. He nonetheless calls the model-and-effort combination the most confusing part of GPT-5.6, and guesses that “Sol on Medium” may become the new default in place of “5.5 xhigh.” Osmani reads it similarly: Sol is not as sharp as Fable but is highly capable, and it resolves much of 5.5’s practical friction around following intent and orchestrating sub-agents. He also notes that it does not quietly degrade output quality when budgets get tight.

Nathan Lambert rejects consolidation onto one model from another angle. Capabilities are jagged, so narrowing to a single model guarantees losing some of them. He uses GPT models for information gathering and research, Claude Opus models for code questions and quick visualization, and Gemini for conceptual explanation and multimodal work.

Armin Ronacher supplies the counterweight. A hands-off harness built on Fable can now work a problem for over thirty minutes without human intervention. And yet, in his words, current hands-off harnesses are producing worse code than what he was building last autumn. Removing human involvement in proportion to the model’s strength lowered the quality of the artifact. It points at the same place as OpenAI’s System Card requiring active supervision on long task sequences.

How to Decide

1. Fill in the free wins first. Measure the cache hit rate. If cache_read_input_tokens is repeatedly zero, something in the prefix (a timestamp in the system prompt, a tool list that varies) is invalidating it. Move anything nobody is waiting on to the Batch API.
2. Sweep effort. Hold the model fixed and vary only output_config.effort (Anthropic) or reasoning.effort (OpenAI), with everything else byte-identical. Include a hard case you know about. A flat curve means the lower setting does this workload’s work.
3. Raise effort only on failures. Where tests or a checker exist, try running everything at low and re-running the failures.
4. Move the model exactly one notch. After stepping down, confirm which effort levels that tier supports and re-sweep from that tier’s own default. Do not carry the previous tier’s setting across.
5. Mix models only when there is bulk. Many independent pieces that no single context window could hold justify an orchestrator. One dependent chain does not.
6. Compare on the tail. On the median task every model looks alike and the cheapest looks best. The bill is set by the hardest tenth: in one 20-problem research run, two problems carried 43% of the spend.

Shapes to Avoid

• Choosing a model from the price list. The unit is cost per completed task. A cheap model’s failure bills its tokens, the retry, and the downstream rework.
• Skipping effort and dropping the model. Lower effort on the newest models often beats the previous generation at high effort. Measure the expensive model at reduced effort before moving down.
• Changing effort mid-conversation. It invalidates the message cache. Test each setting in its own session.
• Splitting the cache by mixing models. Caches are model-scoped, so a cascade forfeits cache reuse across its models.
• Deciding on a single case. A one-task or few-cent difference is inside the noise on single runs. Repeat trials at the settings in contention.
• Switching without an eval. Public benchmarks are saturating and break under paraphrase. A frozen set of 20 to 30 real requests is more useful for the decision.
• Running a strong model unattended. Ronacher’s observation and OpenAI’s System Card point the same way. That autonomous runtime got longer is not the same as supervision becoming optional.

How to Read the Confidence Levels

This note keeps three tiers of evidence separate.

• T1v, vendor first-party: what Anthropic, OpenAI and Google state officially about their own products (documentation, model cards, system cards, pricing pages). Treated as the authority on what a model does, how to use it, its constraints and its price. Claims of competitive superiority made without methodology inside those same documents are separated into the corpus file’s marketing-claim section (Mythos 5’s advantage at vulnerability discovery, Terra’s performance-parity claim, Sol’s Terminal-Bench SOTA claim, Fast mode’s 2.5x).
• T2, public bodies and research firms: UK AISI, NIST CAISI, the International AI Safety Report, the IEA, the European Commission, plus quantitative findings from research firms where the methodology is disclosed. VC-published surveys (Menlo Ventures, a16z) are marked at least partial for investment interest.
• T3, individual opinion: treated as the view of a verifiable technical authority, never quoted as settled fact. Verifiable factual claims carry a “primary source unverified” marker.

Anthropic’s measured figures (the effort curves, re-running failures, the orchestrator’s 55%, Haiku 4.5’s 63% against 92%) come from its own benchmarks, some of which are unnamed. They are useful for direction and do not replace measuring on your own workload.

Related notes: fable-model-tiering-patterns, frontier-model-premium-history-debate.

Unverified Items

• openai.com/index/gpt-5-6/ and openai.com/index/previewing-ultrafast/ return 403, so Sol’s release date, Ultra mode and Ultrafast mode rest on search-result summaries rather than the primary pages.
• The reasoning.mode standard / pro distinction could not be confirmed on the page itself.
• The 280-fold inference-cost drop and the SWE-bench Verified 40%-to-60% figure were taken from secondary aggregation; the arXiv abstract page was reached but the PDF body was not.
• The IEA energy-efficiency figures could not be fetched from the IEA site (403). Several independent articles quote identical numbers.
• Both Gartner predictions (three times more small models, 90% inference-cost drop) have inaccessible primary pages and undisclosed methodology.
• McKinsey’s 60% response-refinement figure was confirmed only from search snippets after repeated fetch timeouts.
• The claims that Terra matches GPT-5.5’s performance and that Sol reaches SOTA on Terminal-Bench 2.1 have no reachable primary description of measurement conditions or evaluation sets. They are recorded as vendor claims.
• No primary source from a public body addressing the price-performance Pareto frontier was found within this collection’s scope. That question belongs to independent evaluation organizations (LMArena, Artificial Analysis, Epoch AI).

References

All accessed 2026-08-29.

Vendor first-party (T1v)

• Anthropic. Introducing Claude Fable 5 and Claude Mythos 5. https://platform.claude.com/docs/en/models/fable-5/introducing-claude-fable-5-and-claude-mythos-5
• Anthropic. Redeploying Claude Fable 5. https://www.anthropic.com/news/redeploying-fable-5
• Anthropic. Choosing the right model. https://platform.claude.com/docs/en/about-claude/models/choosing-a-model
• Anthropic. Optimizing for cost and intelligence. https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence
• Anthropic. Model migration guide. https://platform.claude.com/docs/en/about-claude/models/migration-guide
• Anthropic. Pricing. https://platform.claude.com/docs/en/about-claude/pricing
• Anthropic. Model deprecations. https://platform.claude.com/docs/en/about-claude/model-deprecations
• Anthropic. Claude Code model configuration. https://code.claude.com/docs/en/model-config
• OpenAI. GPT-5.6: Frontier intelligence that scales with your ambition. https://openai.com/index/gpt-5-6/ [body unreachable, 403]
• OpenAI. Model guidance. https://developers.openai.com/api/docs/guides/latest-model
• OpenAI. GPT-5.6 Sol model page. https://developers.openai.com/api/docs/models/gpt-5.6-sol
• OpenAI. Pricing. https://developers.openai.com/api/docs/pricing
• OpenAI. Prompt caching guide. https://developers.openai.com/api/docs/guides/prompt-caching
• OpenAI. Batch guide. https://developers.openai.com/api/docs/guides/batch
• OpenAI. Deprecations. https://developers.openai.com/api/docs/deprecations
• OpenAI. GPT-5.6 System Card. https://deploymentsafety.openai.com/gpt-5-6
• OpenAI Developer Community. Introducing GPT-5.6 series: Sol, Terra and Luna. https://community.openai.com/t/introducing-gpt-5-6-series-sol-terra-and-luna-coming-july-9-10am-pt/1384931
• Google. Gemini API pricing. https://ai.google.dev/gemini-api/docs/pricing

Public bodies and research firms (T2)

• UK AI Security Institute. Frontier AI Trends Report. 2025-12-18. https://aisi.gov.uk/frontier-ai-trends-report
• International AI Safety Report: First Key Update. 2025-10-15. https://arxiv.org/abs/2510.13653
• International AI Safety Report 2026. 2026-02-24. https://arxiv.org/abs/2602.21012
• NIST CAISI. Evaluation of DeepSeek AI Models. 2025-09-30. https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks
• NIST CAISI. Evaluation of DeepSeek V4 Pro. 2026-05-01. https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
• IEA. Energy and AI (Executive summary). 2025-04. https://www.iea.org/reports/energy-and-ai/executive-summary
• European Commission. EU AI Act Article 51. https://artificialintelligenceact.eu/article/51/
• Stanford HAI. AI Index Report 2025, Research and Development. https://hai.stanford.edu/ai-index/2025-ai-index-report/research-and-development
• Menlo Ventures. 2025: The State of Generative AI in the Enterprise. 2025-12-09. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
• a16z. The 2025 Enterprise AI Report. 2025-06-10. https://a16z.com/ai-enterprise-2025/
• OpenRouter. State of AI. 2025-12. https://openrouter.ai/state-of-ai
• LangChain. State of Agent Engineering. 2026-06-12. https://www.langchain.com/state-of-agent-engineering
• Gartner. Predicts by 2027 organizations will use small task-specific AI models three times more than general purpose LLMs. 2025-04-09. https://www.gartner.com/en/newsroom/press-releases/2025-04-09-gartner-predicts-by-2027-organizations-will-use-small-task-specific-ai-models-three-times-more-than-general-purpose-large-language-models [403; confirmed through secondary reporting]
• CIO Dive. AI inference costs to drop by 2030, Gartner says. 2026-03-25. https://www.ciodive.com/news/ai-inference-costs-drop-2030-gartner/815725/
• McKinsey QuantumBlack. Cost versus value: Managing agentic AI system performance. 2026-07. https://www.mckinsey.com/capabilities/quantumblack/our-insights/cost-versus-value-managing-agentic-ai-system-performance [summary only]

Individual opinion (T3)

• Willison, S. Claude Fable 5. 2026-06-09. https://simonwillison.net/2026/Jun/9/claude-fable-5/
• Willison, S. GPT-5.6. 2026-07-09. https://simonwillison.net/2026/Jul/9/gpt-5-6/
• Willison, S. Judgement. 2026-07-03. https://simonwillison.net/2026/Jul/3/judgement
• Osmani, A. GPT-5.6 Sol. 2026-07-09. https://addyosmani.com/notes/gpt-5-6-sol/
• Ronacher, A. The Coming Loop. 2026-06-23. https://lucumr.pocoo.org/2026/6/23/the-coming-loop/
• Lambert, N. Use multiple models. 2026-01-11. https://www.interconnects.ai/p/use-multiple-models
• Rizwan, S. OpenAI’s GPT-5.1. 2025-11-13. https://cline.bot/blog/openais-gpt-5-1
• Karpathy, A. Post on generating a story world with Opus 5 and Three.js. 2026-08-02. https://explainx.ai/blog/karpathy-opus-5-lord-of-the-rings-threejs-world-2026 [the original X post is unreachable; a secondary article quoting it in full]