What you will be able to do
- Name the factors Anthropic says to weigh when choosing a Claude model, including effort as a lever inside a single model
- Compare Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5 on latency, price, thinking, default effort and context window
- Choose between the cost-first (Haiku-first) and capability-first (Opus-first) starting paths for a given workload
Key concept
Cost per completed task — The fair way to compare Claude models is by what it costs to get a task done correctly, not by the price per token. A model with a higher token price can still cost less per result if it needs fewer turns, less thinking and fewer retries.
1.The trade-offs behind every model choice
Anthropic's model-selection guide starts with one sentence that frames this whole topic: choosing a Claude model "means balancing capabilities, speed, and cost." It names the questions behind each factor. Capabilities: which features does the application actually need? Speed: how fast must the model respond? Cost: what is the budget for development and for production? Some models add a speed option of their own. Claude Opus 5.5, Opus 5 and Opus 4.8 support fast mode (research preview), which gives up to 2.5x higher output speed at premium pricing.
The guide also adds a fourth factor that candidates tend to miss: effort. Several Claude models accept an effort parameter that trades intelligence against latency and cost inside one model. The guide is direct about it: "Tuning effort is often a better lever than switching models." So model selection is two decisions, not one. You pick a model tier, and then you pick an effort level on that model. Both decisions sit on the same cost-against-intelligence curve.
Anthropic's cost guide makes the same split. Some savings, such as prompt caching and batch processing, cut spend without touching quality. Model choice is not one of them. It belongs to the levers that trade cost against quality, alongside effort, output caps and multi-model architectures. That is why the choice needs measuring and can't simply be looked up. Anthropic's blog explains why the tiers differ in the first place. Every Claude model is trained for coding, agentic work and knowledge work. What changes between tiers is how hard a problem each one can reliably carry, and what that capability costs in price and speed.
Checkpoint 1 of 4· Check yourself
A team's Opus-class deployment meets its quality bar but costs too much. According to Anthropic's model-selection guidance, what should they usually try before switching to a different model?
Effort trades intelligence against latency and cost within one model, and the guide says tuning it is often a better lever than changing models. Fast mode raises output speed at premium pricing, so it doesn't lower cost.
“Tuning effort is often a better lever than switching models.”Source: platform.claude.com
2.The current lineup side by side
The models overview lists four current models. All of them support text and image input, text output, multilingual use, vision and tool use. That means the choice rarely depends on whether a feature exists at all. It depends on latency, price, thinking mode and limits. The table below collects the figures most likely to decide a selection question.
| Model (API ID) | Comparative latency | Price (input / output) | Thinking | Default effort | Context window / max output |
|---|---|---|---|---|---|
| Claude Fable 5.1 (claude-fable-5-1) | Slower | $10 / input MTok, $50 / output MTok | Adaptive (always on) | high | 1M tokens / 128K tokens |
| Claude Opus 5.5 (claude-opus-5-5) | Moderate | $4 / input MTok, $20 / output MTok | Adaptive (always on) | medium | 1M tokens / 128K tokens |
| Claude Sonnet 5.5 (claude-sonnet-5-5) | Fast | $2 / input MTok, $10 / output MTok | Adaptive | high | 1M tokens / 128K tokens |
| Claude Haiku 4.5 (claude-haiku-4-5-20251001) | Fastest | $1 / input MTok, $5 / output MTok | Extended | Not supported | 200K tokens / 64K tokens |
Three rows in the table catch people out. First, Haiku 4.5 is the odd one out: it has a 200K-token context window rather than 1M, it uses extended thinking rather than adaptive thinking, and the effort parameter is listed as not supported. If a workload needs to fit very long inputs in one context, or depends on effort tuning, Haiku is ruled out before price comes into it. Second, Opus 5.5 defaults to medium effort, while the other effort-capable models default to high. Third, Fable 5.1 is described as Anthropic's most capable model open to all customers. Its sibling Claude Mythos 5.1 has the same capabilities but is available only to Project Glasswing participants, so access can limit the choice too.
The selection guide matches each model to a need. Fable 5.1 is for the highest available capability, such as agent sessions that run for hours and multistep deep research. Opus 5.5 is for complex agentic coding and enterprise work. Sonnet 5.5 offers speed and capability for everyday coding, agent and enterprise workloads. Haiku 4.5 offers the lowest latency and price and still has extended thinking, which suits real-time applications and sub-agent tasks.
Checkpoint 2 of 4· Match them up
Match each model to the need Anthropic's selection guide pairs it with
Tap a term, then the definition that fits it.
These pairings come straight from the 'When you need... / Consider starting with...' table in the choosing-a-model guide.
“The lowest latency and price, with extended thinking”Source: platform.claude.com
3.Two ways to pick a starting model
The selection guide describes two ways to begin, and an exam scenario usually points clearly to one of them. The cost-first path starts small. Begin implementation with Claude Haiku 4.5, test the use case thoroughly, check whether performance meets your requirements, and "Upgrade only if necessary for specific capability gaps." The guide recommends this path for initial prototyping, tight latency requirements, cost-sensitive implementations, and high-volume, straightforward tasks.
The capability-first path starts at the top. Implement with Claude Opus 5.5, optimise your prompts for that model, and evaluate. Then "Consider increasing efficiency by lowering effort or downgrading models over time with greater workflow optimization." If evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, move up to Claude Fable 5.1. This path suits complex reasoning, scientific or mathematical applications, tasks that need nuanced understanding, cases where accuracy matters more than cost, and advanced coding and high-autonomy agentic work.
You may also see an older rule of thumb in Anthropic's pricing guidance: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning. The roles of the tiers haven't changed. What has changed is the recommended default. Current guidance says "Most workloads start with Claude Opus 5.5," and Anthropic's blog explains why. More capable models often finish a task in fewer turns, and starting with a smaller model can make it harder to tell model failures apart from setup failures.
Checkpoint 3 of 4· Put it in order
Put the steps of the capability-first approach in the order the selection guide gives them
- 1.Implement with Claude Opus 5.5
- 2.Consider lowering effort or downgrading models over time
- 3.Evaluate if performance meets your requirements
- 4.If evals at xhigh or max effort still fall short, move to Claude Fable 5.1
- 5.Optimize your prompts for this model
The guide sets this sequence: start at the strongest practical starting point, tune prompts, evaluate, then optimise for efficiency. It escalates to Fable 5.1 only when the highest effort levels on Opus 5.5 aren't enough.
“Consider increasing efficiency by lowering effort or downgrading models over time with greater workflow optimization.”Source: platform.claude.com
Checkpoint 4 of 4· Exam question
A data engineering team needs to classify and summarize 40 million historical support tickets overnight. There is no interactive user waiting on responses, and the team wants to minimize total API spend while keeping quality reasonably high. Which combination of choices best reflects the intended cost trade-off for this workload?
Correct answer: A — Submit the requests through the Message Batches API, which processes large asynchronous volumes at roughly half the cost of standard synchronous API calls
- A. Correct. Batch processing lets you process large volumes of requests asynchronously for cost savings, with Batch API calls costing 50% less than standard API calls, which directly matches a non-interactive, cost-minimizing overnight workload.
- B. Incorrect. Prompt caching reduces cost and latency when repeated context is reused; disabling it does not itself reduce charges and would forgo available savings on any repeated instructions or context across the ticket batch.
- C. Incorrect. Batch API discounts apply broadly across models, not exclusively to the highest-tier model, and choosing Opus 4.8 for simple classification would add unnecessary per-token cost relative to a lighter model.
- D. Incorrect. Even without latency constraints, maximizing effort on every request increases token usage and cost without necessarily improving results on straightforward classification and summarization tasks, working against the stated goal of minimizing spend.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Every current Claude model supports the effort parameter and a 1M-token context window, so they differ only in price.Why is that wrong?
Claude Haiku 4.5 lists effort as not supported, uses extended rather than adaptive thinking, and has a 200K-token context window. Those limits can rule it out before price is even considered.
Covered in The current lineup side by side
2.Anthropic always recommends starting with the cheapest model and working upward.Why is that wrong?
Starting with Haiku is one of two documented paths, meant for prototyping, tight latency, cost sensitivity and high-volume simple tasks. For most workloads the current default starting point is Opus 5.5.
Covered in Two ways to pick a starting model
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Choosing a Claude model means balancing capabilities, speed, and cost.”
↩︎ The trade-offs behind every model choice“fast mode (research preview), which delivers up to 2.5x higher output speed at premium pricing.”
↩︎ The trade-offs behind every model choice“Claude Fable 5.1 (claude-fable-5-1) is Anthropic's most capable model open to all customers.”
↩︎ The current lineup side by side“Upgrade only if necessary for specific capability gaps.”
↩︎ Two ways to pick a starting model“If your evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, move to Claude Fable 5.1.”
↩︎ Two ways to pick a starting model“Most workloads start with Claude Opus 5.5.”
↩︎ Exam trap 2“Tuning effort is often a better lever than switching models.”
↩︎ Checkpoint“The lowest latency and price, with extended thinking”
↩︎ Checkpoint“Consider increasing efficiency by lowering effort or downgrading models over time with greater workflow optimization.”
↩︎ Checkpoint - 2.https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligenceOfficial docs
“Tradeoffs exchange cost for intelligence: model choice, effort, output caps and task budgets, an elapsed-time clock, and multi-model architectures.”
↩︎ The trade-offs behind every model choice“Compare on cost per completed task, not per token”
↩︎ Key concept - 3.https://claude.com/blog/claude-models-explained-choosing-the-best-model-for-your-use-caseSecondary source
“how hard a problem they can reliably carry, and what that capability costs in price and speed”
↩︎ The trade-offs behind every model choice“Starting with a smaller model can also make it harder to distinguish between model failures and setup failures.”
↩︎ Two ways to pick a starting model - 4.https://platform.claude.com/docs/en/models/overviewOfficial docs
“All current models support text and image input, text output, multilingual capabilities, vision, and tool use.”
↩︎ The current lineup side by side“Context window | 1M tokens | 1M tokens | 1M tokens | 200K tokens”
↩︎ The current lineup side by side“Default effort | high | medium | high | Not supported”
↩︎ Exam trap 1“If you're unsure which model to use, start with Claude Opus 5.5 for most workloads.”
↩︎ Prediction - 5.
“Choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning”
↩︎ Two ways to pick a starting model