Claude Sonnet 5 Model Selection: The 5-Minute Developer Guide
By EndOfCoding
As of July 1, 2026, Claude Sonnet 5 is the new default model on Free and Pro plans. Meanwhile, Fable 5 left subscriptions on July 7 — it is now API-only and billed separately. Together these two shifts create a concrete decision point for every developer: which model should you actually be using for which task, and what does your CLAUDE.md need to say right now? This quick guide makes that decision in under 5 minutes. If you have Fable 5 hardcoded anywhere in your config, you are already paying API rates you may not have budgeted for. Read on.
What You'll Learn
You will learn the current four-model lineup — Haiku 4.5, Sonnet 5, Opus 4.8, and Fable 5 — including benchmarks and pricing for each. You will see a decision table matching task types to the right model so you stop over-spending on Opus for routine work. You will get a minimal CLAUDE.md model-routing block you can paste in today. You will understand how Sonnet 5's self-verification feature changes your code review workflow. And you will know the right monthly budget targets based on your usage level before and after the Sonnet 5 promotional pricing window closes on August 31.
Step 1: Know the Current Lineup
The four-model stack as of July 2026:
| Model | Best For | Input / Output (per 1M tokens) |
|---|---|---|
| Haiku 4.5 | Batch jobs, CI/CD, simple lookups | $0.25 / $1.25 |
| Sonnet 5 | Most coding tasks, default for Pro | $2 / $10 (promo until Aug 31) |
| Opus 4.8 | Complex architecture, hard bugs | $5 / $25 |
| Fable 5 | Top-ceiling tasks only (API-only) | $10 / $50 |
Sonnet 5 benchmarks within 0.5 percentage points of Opus 4.8 on HLE agentic coding tasks. Fable 5 is meaningfully better only for the top 5% of tasks that genuinely hit the performance ceiling.
Step 2: Match Your Task to the Right Model
Use this decision table before every session:
| Task | Right Model | Why |
|---|---|---|
| CRUD features, tests, docs | Sonnet 5 | More than fast enough; saves 60%+ over Opus |
| Refactors, integrations, PR review | Sonnet 5 | Self-verification handles edge cases |
| Complex system architecture | Opus 4.8 | Needs deeper multi-step reasoning |
| Security audit on unfamiliar codebase | Opus 4.8 | Context + judgment depth matters |
| CI/CD file scanning (1000s of files) | Haiku 4.5 | Volume task; quality ceiling is irrelevant |
| Batch content generation | Haiku 4.5 | Cheapest per token at scale |
| Novel research, extreme reasoning | Fable 5 (API key required) | Only when you have a specific benchmark need |
For the full cost-optimization framework across all your AI tools, see the AI Coding Cost Audit on EndOfCoding.
Step 3: Add Model Routing to Your CLAUDE.md
Paste this block into your project CLAUDE.md and adjust the task list to your stack:
## Model Selection Policy
- Default: claude-sonnet-5 (all coding tasks, PR review, refactors, documentation)
- Use claude-opus-4-8 ONLY for: system architecture, hard cross-service bugs, security audits on unfamiliar codebases
- Use claude-haiku-4-5 ONLY for: batch CI scans, formatting passes, simple lookups, background routines
- DO NOT use claude-fable-5 unless a specific benchmark comparison requires it and you have a separate API key budgeted for it
If you previously had claude-fable-5 or fable hardcoded in your config, change it to claude-sonnet-5 now. Fable 5 is no longer included in any subscription — every call bills at $10/$50 per million tokens.
For Claude engineers in demand on job platforms, see the model routing requirements in current JDs at LLMHire.
Step 4: Use Sonnet 5's Self-Verification for Code Review
Sonnet 5 ships with a self-verification capability: after generating code, it can re-examine its own output against a set of constraints you provide before returning a response. This changes how you structure code review prompts.
Instead of:
"Write a login route with rate limiting."
Use:
"Write a login route with rate limiting. Before returning, verify: (1) no hardcoded secrets, (2) input validated before DB call, (3) rate limit applied before auth check, (4) error messages do not leak user existence."
Sonnet 5 will flag its own issues in a verification block before the final output. You catch problems at generation time rather than review time. This makes Sonnet 5 meaningfully more useful for security-sensitive code than earlier Sonnet versions — even without switching to Opus.
The Vibe Coding Ebook tools landscape chapter has been updated with Sonnet 5's self-verification syntax.
Step 5: Set a Monthly Budget
Concrete targets based on usage level, assuming Sonnet 5 as your primary model:
| Usage Level | Description | Monthly Target |
|---|---|---|
| Light | 2-3 hours/day, mostly Sonnet 5 | $20-35 |
| Moderate | 4-5 hours/day, occasional Opus | $50-80 |
| Heavy | 6+ hours/day, background agents | $90-150 |
| Power | Multiple parallel agents, CI pipelines | $150-250 |
Note: After August 31, Sonnet 5 moves from $2/$10 to $3/$15 per million tokens. That is still 40% cheaper than Opus 4.8 at $5/$25 — so the routing logic above stays the same. Budget targets increase roughly 20-30% if you are Sonnet-heavy. Set a billing alert at 80% of your cap in the Anthropic Console today.
Common Challenges
'Will Sonnet 5 produce lower quality code than Fable 5?' — Sonnet 5 scores within 0.5 percentage points of Opus 4.8 on HLE agentic coding benchmarks. Fable 5 is only meaningfully better for the top 5% of tasks that genuinely hit the performance ceiling — novel research, extreme multi-step reasoning, tasks you cannot solve with any other model. For 85-90% of everyday developer work, Sonnet 5 is indistinguishable from Fable 5 in practice. The quality drop you are worried about is not measurable on routine code. 'I have Fable 5 hardcoded in my CLAUDE.md.' — Change it to claude-sonnet-5 right now. Fable 5 exited subscriptions on July 7. Every call you make with a Fable 5 model ID while on a Pro or Max plan bills at API rates: $10 per million input tokens, $50 per million output tokens. A single 3-hour agentic session can generate 2-5 million tokens. That is a $20-50 surprise charge per session on top of your subscription. Check your config files today: CLAUDE.md, .cursorrules, any agent scripts or scheduled routines that specify a model ID. 'What happens when the Sonnet 5 promotional pricing ends August 31?' — After August 31, Sonnet 5 moves from $2/$10 to $3/$15 per million tokens. It is still 40% cheaper than Opus 4.8 ($5/$25) and 70% cheaper than Fable 5 ($10/$50). The model routing decision — Sonnet 5 for most tasks, Opus 4.8 for the hard 10-15% — does not change after promotional pricing ends. Your budget targets will increase by 20-30% if you are Sonnet-heavy. Build that into your August planning now.
Advanced Tips
Enable Sonnet 5 adaptive thinking in CLAUDE.md for complex tasks. Sonnet 5 supports an adaptive thinking mode that lets it reason deeper on hard problems without requiring a model switch to Opus. Add this to your CLAUDE.md: For tasks involving cross-service data flows, race conditions, or security policy design, use extended thinking before responding. Sonnet 5 will increase its reasoning depth automatically for those task types, closing most of the gap with Opus 4.8 at no extra model-switch overhead. Use GLM-5.2 for high-volume CI/CD batch scans. GLM-5.2 (MIT license, self-hostable, free at scale) is a strong option for the batch scanning tier — the 1,000-file security scans, the lint passes, the large-scale code search tasks that run thousands of requests per day. At that volume, even Haiku 4.5 at $0.25/$1.25 adds up. GLM-5.2 on your own infrastructure is effectively $0 per token beyond compute. Set it as the model for your background CI routines and reserve Haiku for tasks that need the Anthropic API's reliability SLA. Cross-reference the cost audit post for the full spending framework. The AI Coding Cost Audit on this blog covers the three optimization levers in detail: model routing, prompt caching, and task scoping. Implementing all three together produces the 40-60% cost reduction most developers are targeting. This post handles model routing; that post handles the rest.
Conclusion
Sonnet 5 handles 85-90% of developer tasks at a fraction of Opus 4.8 cost. Reserve Opus 4.8 for the 10-15% of work that genuinely needs it: complex architecture decisions, hard cross-service bugs, security audits on unfamiliar codebases. Do not use Fable 5 unless you have a specific benchmark requirement, a separate API budget allocated for it, and a task that demonstrably exceeds Sonnet 5's ceiling. The model lineup is more capable and more cost-differentiated than it was six months ago. The developers who route correctly will spend 40-60% less than those defaulting to the top model for everything. Update your CLAUDE.md today. Check your config files for any Fable 5 model IDs. Set a billing alert at 80% of your monthly target. For hands-on model selection exercises, the Vibe Coding Academy courses walk through task classification and routing decisions with real codebases. For the latest model releases and pricing changes as they happen, subscribe to the EndOfCoding newsletter.
Build Blueprint · Creator
Have an idea? Get the spec your AI agent can build from.
Describe any product and get a complete build blueprint — stack, data model, screens, APIs, and a ready-to-paste prompt for Claude Code or Cursor. Export to PDF.
Open the Blueprint ▸