Which Claude model to use, and why — the structure behind the lineup, so the logic still holds after the specific names and numbers change.
Prepared by Tienta, Inc. · Last reviewed August 2026
Start here
Anthropic ships new models every few months and retires old ones on a rolling basis. Rather than pin exact specs here — they would be stale within weeks — this guide explains how the model lineup is structured, so the logic holds even after the specific names and numbers change, and links to the two pages Anthropic itself keeps current. Check those two first if a number below looks off.
Part 1
This is the part that doesn't change often. Claude models come in size tiers, named after forms of writing, from fastest and cheapest to most capable.
| Tier | Named for | General role |
|---|---|---|
| Haiku | Short form | Fastest, cheapest, highest-volume tasks |
| Sonnet | Mid-length form | The default workhorse — best balance of speed, cost, and intelligence |
| Opus | Long-form, most capable | Complex reasoning, coding, high-stakes work |
| Fable | Newer top tier | Anthropic's most capable widely released model, above Opus |
Each tier gets versioned over time — a tier might go through several numbered releases in a year. The tier name tells you the role; the version number tells you how recent it is. When a client asks “should we use the new one,” the real question is usually “which tier does our task need,” not “which is newest.” A newer Haiku is not a substitute for Opus-level reasoning; it's just a faster, cheaper Haiku.
There is also a limited-access tier, Mythos, available only to vetted organizations for specific sensitive use cases such as defensive cybersecurity, through Anthropic's invitation-only Project Glasswing program. Not something most clients will encounter or need.
Part 2
Pull the live numbers from the Models overview table before sending anything to a client — pricing and context windows shift with each release.
| Tier | Best for | Context window | Relative cost |
|---|---|---|---|
| Haiku | Real-time, high-volume, cost-sensitive tasks; simple classification, tagging, routing, first-pass drafts | Smaller | Lowest |
| Sonnet | The default choice for most production work — coding, data analysis, content creation, agentic tool use | Largest tier available | Mid |
| Opus | Complex agentic coding, enterprise work, multi-hour autonomous agents, large-scale refactoring, vision-heavy workflows | Largest tier available | High |
| Fable | The highest available capability — long-running agents, deepest reasoning, long-horizon agentic tasks, advanced research | Largest tier available | Highest |
Anthropic's own rule of thumb, verbatim from their docs: “If you're unsure which model to use, start with Claude Opus [current flagship] for complex agentic coding and enterprise work. For workloads that need the highest available capability, use Claude Fable [current version].”
Part 3
Anthropic's framework offers two entry points, and they suit different situations.
Begin with the cheapest and fastest tier, test the real use case, and upgrade only where there's a demonstrated capability gap. Best for prototyping, tight-latency applications, and cost-sensitive or high-volume straightforward tasks.
Begin with the top tier, get it working well, then step down to a cheaper tier once the workflow is proven and optimized. Best for complex reasoning, scientific and mathematical work, advanced coding, and high-autonomy agentic work — anywhere accuracy matters more than early-stage cost.
Several current models support an effort parameter that trades intelligence for speed and cost within the same model — often a better first move than switching tiers entirely. If a workflow feels slow or expensive on a capable model, try lowering effort before downgrading the model.
Route bulk and simple work to a cheap tier and escalate only the hard decisions to a frontier model, so most tokens are billed at the lower rate. Two common patterns:
Part 4
This connects back to the Admin Cheat Sheet — model selection shows up differently depending on which product a client is using.
| Product | Where model choice lives | Who controls it |
|---|---|---|
| claude.ai chat | Model picker in the chat UI | The individual user, per conversation. Every paid plan can access all current tiers — the plan controls usage volume, not which models are available. |
| Claude Code (Team/Enterprise) | /model in the CLI, or an org-wide default set by an admin | Org admins can set a default model for Claude Code users at claude.ai/admin-settings/claude-code, org-wide or per custom role. |
| Claude API (platform.claude.com) | The model field in each API request, referencing a specific model ID such as claude-opus-5 | The developer, per API call. This is the one place model IDs matter to the byte, since a pinned ID is what gets billed and what determines behavior. |
One nuance worth flagging to clients: on the API, model IDs are pinned snapshots, not evergreen pointers. claude-opus-5 won't silently start meaning a different model next year the way “the current flagship” might in conversation. That's a feature for production stability, but it also means client integrations need a deliberate migration step when a new tier version ships — Anthropic publishes a migration guide for exactly this.
Part 5
| Client says… | Point them to… |
|---|---|
| “Which model should we default to?” | The Sonnet tier, for nearly everyone — it's Anthropic's own “best combination of speed and intelligence” framing. Reserve Opus and Fable for specific hard tasks, not as the default. |
| “Our API bill is high — can we use a cheaper model?” | Try lowering the effort parameter first, then consider a mixed executor/advisor setup before a blanket downgrade — see §3.4. |
| “Do we need to keep upgrading to the newest model?” | Not automatically — pinned model IDs mean nothing breaks by staying put. But check Model deprecations for retirement dates. |
| “What's the difference between claude.ai and the API?” | Same models; a different mechanism for choosing them (picker versus explicit ID) and different billing (per seat versus per token). See the Admin Cheat Sheet. |
| “Can non-technical staff pick the ‘best’ model themselves?” | Yes, in the claude.ai chat picker — every paid plan includes all current tiers. Consider an organization instruction reinforcing “Sonnet is the default, Opus is for X” so people aren't guessing per conversation. |
Part 6
Bookmark these rather than quoting numbers from them — the pages stay current on their own.
Tienta sets up and administers Claude for Team and Enterprise clients — roles and groups, SSO and SCIM, spend controls, and the Admin API plumbing behind them.
