Claude Model Comparison 2026: Opus vs Sonnet vs Haiku vs Fable

Claude Fable 5 scores 80.3% on SWE-bench Pro, 11 points above Opus 4.8's 69.2%, at exactly twice the per-token cost. Anthropic's model documentation notes that the performance gap narrows on shorter, well-scoped tasks. With four active model tiers spanning a 10x price range, the real question isn't which Claude model is best. It's which one earns its price for your specific workload.

Anthropic currently offers four Claude model tiers, each targeting a different point on the speed-cost-capability spectrum. The model you choose also matters for Claude Cowork and Claude Code, where autonomous agent sessions consume more tokens than standard chat.

Fable 5 is the most capable model Anthropic has released publicly. It carries always-on adaptive thinking, a 1M token context window, and 128k token max output. At $10 input and $50 output per million tokens, it costs twice as much as Opus and is built for the longest, hardest problems. Anthropic released it on June 9, 2026.

Opus 4.8 is the flagship of the Opus tier. Same 1M context window and 128k output ceiling, but at $5/$25 per million tokens. It supports adaptive thinking and an effort parameter that lets you dial reasoning depth up or down without switching models. Anthropic positions it for complex reasoning, agentic coding, and high-autonomy work.

Sonnet 4.6 splits the difference between speed and intelligence. It shares the 1M context window with the larger models and offers both extended thinking and adaptive thinking. At $3/$15 per million tokens, it handles code generation, data analysis, and content creation at roughly 60% of Opus pricing.

Haiku 4.5 is the speed tier. A 200k context window and 64k max output keep it compact. At $1/$5 per million tokens, it's 5x cheaper than Opus on both input and output. Anthropic describes it as "near-frontier intelligence" and recommends it for real-time applications, high-volume processing, and sub-agent tasks where latency matters more than reasoning depth.

Every model accepts text and image input, supports multilingual tasks, and is available through the Claude API, AWS Bedrock, Google Cloud, and Microsoft Foundry.

This table covers the current Claude model lineup as of June 2026, using data from Anthropic's platform documentation.

A few details worth noting. Fable 5 and Opus 4.8 both offer 1M context windows, but their tokenizers differ. Opus 4.7 introduced a new tokenizer that Fable 5 shares, and it produces roughly 30% more tokens from the same text compared to earlier models. Your effective context window in words is somewhat smaller with these newer tokenizers.

Haiku 4.5 is the only current model without adaptive thinking and the only one with a 200k context window instead of 1M. Those constraints are part of the speed tradeoff. For tasks that fit within 200k tokens and don't need deep reasoning chains, Haiku delivers answers faster than any other Claude model.

Adaptive Thinking vs Extended Thinking

These are two different reasoning mechanisms. Extended thinking gives the model an explicit scratchpad to work through problems step by step before responding. Adaptive thinking dynamically adjusts reasoning depth based on task complexity, without a separate thinking phase. Fable 5 and Opus 4.8 use adaptive thinking (Fable has it always on). Sonnet 4.6 supports both modes. Haiku 4.5 supports extended thinking but not adaptive thinking.

The Effort Parameter

On Opus 4.8, the effort parameter deserves attention. Anthropic defaults effort to high on all surfaces, but recommends xhigh for coding and agentic tasks. You can also drop effort to medium or low to save tokens on simpler prompts. This within-model tuning often matters more than switching between model tiers entirely.

Model selection is a workload decision, not a quality decision. Every tier produces good output. The question is whether that output justifies the cost for your specific use case.

Long, Autonomous Agent Sessions

Fable 5 earns its premium on tasks that run for hours with minimal human oversight. Multi-file refactoring across large codebases, research that requires synthesizing dozens of sources, and agent workflows that branch into parallel subtasks all benefit from Fable's higher ceiling. On FrontierCode Diamond, which tests exactly this kind of extended autonomous work, Fable 5 scored 29.3% versus Opus 4.8's 13.4%.

If your agent sessions routinely exceed 30 minutes or involve coordinating multiple tool calls across complex state, Fable 5 is the pragmatic choice despite the price.

Complex Coding and Reasoning

Opus 4.8 hits the sweet spot for most development work. It scores 69.2% on SWE-bench Pro and 88.6% on SWE-bench Verified, both strong numbers for production coding tasks. The effort parameter gives you a tuning knob: set xhigh for architecture decisions and tricky bugs, then drop to high for standard feature work.

Opus 4.8 also supports fast mode (a research preview feature) at premium pricing, delivering up to 2.5x higher output speed when latency is critical during coding sessions.

Content Creation and Analysis Sonnet 4.6 handles writing, summarization, and data analysis well at 40% lower cost than Opus. Its extended thinking capability lets it work through multi-step problems when prompted, and the 1M context window means it can process entire codebases or document sets in a single conversation.

For content teams generating articles, reports, or marketing copy, Sonnet typically produces output comparable to Opus at a meaningful cost savings. The difference shows up in edge cases that require deep domain reasoning or extended chains of logic.

High-Volume and Real-Time Applications

Haiku 4.5 is the right choice for three scenarios: customer-facing chatbots where response time matters, batch processing pipelines where you're running thousands of prompts, and sub-agent tasks where a larger model delegates simpler work to faster workers.

At $1/$5 per million tokens, you can process 5x more volume than Opus for the same budget. For classification, extraction, simple Q&A, and routing tasks, Haiku's quality gap compared to Opus is negligible while the cost savings are substantial.

Try the Effort Dial Before Switching Models

Before upgrading from one model tier to another, adjust the effort parameter. On Opus 4.8, bumping from high to xhigh can close the gap with Fable 5 on many tasks at half the per-token price. Similarly, dropping Opus to medium effort for simple tasks may give you Sonnet-quality costs with Opus-quality reasoning available when you need it.

This is Anthropic's recommended approach: tune effort before reaching for a different model.

Give your Claude agents a workspace that persists between sessions

Fast.io includes shared storage, MCP server access, and built-in RAG for searching agent outputs by meaning. Starts with a 14-day free trial.

Start 14-Day Trial

Claude Cowork runs autonomous multi-step tasks on your desktop, and it consumes more of your usage allocation than standard chat. Anthropic's enterprise consumption guide categorizes both Cowork and Claude Code as "higher intensity" surfaces compared to "lower intensity" core chat.

The higher consumption comes from how Cowork operates. Each autonomous session involves screenshots for visual understanding, sub-agent coordination across workstreams, file system operations, and multi-turn reasoning chains. All of that generates intermediate tokens that count against your plan's usage limit even though you only see the final output.

Plan Economics by Model Choice

Your plan determines how much total usage you get. Your model choice determines how fast you burn through it. Anthropic's documentation states that "Opus can consume several times more tokens than Sonnet for the same task."

Pro ($20/month) includes Cowork, Claude Code, and Claude Design. It provides a baseline usage allocation. For developers running a few Cowork sessions per week, Pro works if those sessions use Sonnet or Haiku. Extended Opus sessions will hit the limit faster.

Max 5x ($100/month) gives five times the Pro usage allocation. This is the practical minimum for regular Cowork and Claude Code users who prefer Opus. The 5x multiplier absorbs the higher token consumption of agentic sessions without constant limit-watching.

Max 20x ($200/month) targets power users running long Cowork sessions daily. At 20x Pro's allocation, you can sustain multiple hours of Opus-powered autonomous work per day.

Team Standard ($25/seat/month) bundles SSO, admin controls, and more usage than Pro for organizations.

Team Premium ($125/seat/month) adds 5x more usage than Standard seats and is the Team tier with Claude Code access.

A Practical Routing Pattern

Use Opus or Fable for tasks that genuinely need deep reasoning, and route everything else to Sonnet or Haiku. If you're building agent workflows through the API, you can automate this routing based on task complexity. If you're using Cowork interactively, you can switch models within the interface.

For teams that want to stretch their usage budget further, offloading file storage and retrieval to an external workspace reduces the tokens Claude spends on file management overhead. Fast.io's agent storage provides persistent storage with MCP access, so Claude can read and write files through a workspace API rather than processing everything through screenshots and local file system operations. The workspace also gives you version history and audit trails for anything the agent creates, plus Intelligence Mode for searching outputs by meaning.

Anthropic recommends two starting strategies. Option one: begin with Haiku 4.5 and upgrade only when you hit capability gaps. Option two: start with Opus 4.8, optimize your prompts, then drop to cheaper models where possible. Both approaches require actual testing against your own workloads, not benchmark comparisons from blog posts.

Create a Benchmark Set

Pick 20 to 50 representative tasks from your actual work. Include easy cases (classification, formatting, extraction), medium cases (summarization, code review, content generation), and hard cases (multi-step reasoning, architecture decisions, novel problem solving). Run each task through every model tier and score the outputs on accuracy, quality, and completeness.

The results will almost show that Haiku handles the easy tier well, Sonnet covers the easy and medium tiers, and Opus or Fable only prove necessary for the hard tier. That distribution tells you where to route each task type.

Track Costs Per Task Type

Every API call returns token counts. Log them by task type and model. After a week of data, you'll know exactly what each task category costs across tiers. Common findings: extraction tasks cost 10x less on Haiku than Opus with no quality difference. Complex coding tasks cost 2x more on Fable than Opus but save developer time that more than covers the premium.

For batch workloads, take advantage of the Batch API's 50% discount. Prompt caching charges only 10% of the input price, so repeated system prompts and context become nearly free after the first call.

Manage Agent Output Across Test Runs

When you're testing multiple models against the same prompts, you need a consistent way to store and compare outputs. Local folders work for small tests, but for team evaluations or longer-running agent workflows, a shared workspace keeps everything organized.

Fast.io workspaces give each team member and each agent access to the same files through both the web UI and the MCP server. Enable Intelligence Mode on a workspace to index outputs for semantic search, so you can ask questions across hundreds of test results without reading each file. Plans start with a 14-day free trial, with workspaces, monthly AI credits, and storage scaled to each tier.

Agents can build workspaces, populate them with test outputs, and then transfer ownership to a human reviewer. The human gets full access through the browser while the agent retains admin permissions for future runs.