Claude Sonnet 5 Why Cheaper Tokens Cost You More
You see a new model drop with a 33% price cut and immediately swap out your API keys. A week later, your agent's token usage has exploded, and the output quality actually looks worse.
We've been watching the LLM price wars closely, and the latest evaluation data reveals a massive catch in the "cheaper is better" narrative.
News Summary
Recent benchmarks tested Claude Sonnet 4.6 against the newer Sonnet 5 across 150 agent tasks in VS Code. The workloads covered architecture design and SharePoint Framework upgrades using GitHub Copilot Chat.
On paper, Sonnet 5 looks like an easy win, boasting a 33% per-token price drop across all input and output categories. But rate cards don't pay the bills-total token consumption does.
During architecture tasks, Sonnet 5 consumed 12x more tokens at the median compared to 4.6. One specific run saw the newer model chew through 47x the typical token volume. This aggressive consumption completely destroys the 33% discount on complex tasks.
For code upgrades, Sonnet 5 cost $2.01 per run, making it 3.7x more expensive than Sonnet 4.6's $0.55 average. Quality took a hit on routine tasks, too. On architecture scenarios, the older Sonnet 4.6 produced highly idiomatic, usable output 90% of the time, while Sonnet 5 dropped to 78%.
The newer model isn't entirely useless, though. Sonnet 5 proved highly capable at strict instruction following. When told to upgrade a project to a specific version over what the docs recommended, Sonnet 5 succeeded 100% of the time, whereas 4.6 failed every attempt.
Developer Impact
If you're building an autonomous agent or a tool that relies on zero-shot generation, this token variance is a massive risk. You cannot budget for an API that swings from 16,000 tokens to 6.6 million tokens on the exact same prompt.
For SaaS founders wrapping these APIs, switching your default routing to Sonnet 5 without running evals could wreck your margins overnight. Stick to 4.6 for routine architecture or design tasks where predictability and established patterns matter.
Switch to Sonnet 5 only when your pipeline demands strict adherence to system prompts over retrieved context.
Our Analysis
We're seeing a fundamental shift in how model upgrades behave. The "jagged frontier" of AI capabilities means newer models aren't strictly better-they just fail differently.
Anthropic clearly optimized Sonnet 5 for deeper reasoning and exhaustive search. We see this when the model attempts to dig up undocumented migration steps, sometimes spinning up 69 million tokens in a single run. But effort doesn't equal execution.
The most critical takeaway here is that LLMs have hit a content ceiling, not a capability ceiling. Neither model could execute structural toolchain migrations because the steps simply aren't documented. You can't compute your way out of a knowledge gap.
We predict the next wave of developer tools will focus heavily on agent extensions and dynamic context retrieval rather than raw model routing. Fixing your RAG pipeline or injecting better documentation into your context window will yield a higher ROI than chasing the next point-release model upgrade.
| Metric | Claude Sonnet 4.6 PDF | Claude Sonnet 5 PDF |
| Input Price (per 1M) | $3.00 | $2.00 |
| Output Price (per 1M) | $15.00 | $10.00 |
| Avg. Cost (Code Upgrade) | $0.55 | $2.01 |
| Task Completion (Upgrades) | 60% | 100% |
| Output Quality (Architecture) | 90% | 78% |
To handle this variance, implement a routing layer that sends strict, context-heavy tasks to Sonnet 5, while keeping routine generation on 4.6.
// Minimal model router based on task constraints
import { Anthropic } from '@anthropic-ai/sdk';
const anthropic = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
async function routeAgentTask(prompt, requiresStrictAdherence) {
// Use Sonnet 5 only if we need aggressive instruction following
const modelId = requiresStrictAdherence
? 'claude-3-5-sonnet-latest'
: 'claude-3-haiku-20240307'; // Fallback to cheaper, predictable models
// Set strict max_tokens to prevent the 6.6M token runaway loops
const maxTokens = requiresStrictAdherence ? 8000 : 4000;
try {
const response = await anthropic.messages.create({
model: modelId,
max_tokens: maxTokens,
messages: [{ role: 'user', content: prompt }],
});
return response.content;
} catch (error) {
console.error("Routing failed:", error);
throw error;
}
}FAQs
Q: Is Claude Sonnet 5 cheaper than 4.6?
A: On paper, Sonnet 5 is 33% cheaper per token. In practice, it consumes significantly more tokens per prompt, often making it more expensive for complex tasks.
Q: Why does Sonnet 5 use so many tokens?
A: The newer model tends to perform exhaustive web fetching and deeper searching on complex tasks. This leads to wild variance, with runs swinging from 16K to 6.6M tokens on the same prompt.
Q: Which model should I use for coding tasks?
A: Use Sonnet 5 for code upgrades where you need strict adherence to specific instructions. Stick to 4.6 for routine architecture tasks where predictability and idiomatic output are prioritized.
Q: Can newer models fix bad documentation?
A: No, an agent can only use the information it finds. If migration steps or toolchain configurations aren't documented, a newer model won't solve the problem.
Our Take
Stop blindly upgrading your default models just because the API provider announced a price cut. Measure your specific workloads, track your token variance, and fix your internal documentation first. We'll be keeping a close eye on how Anthropic handles this token bloat in upcoming point releases-because right now, predictability is more valuable than a theoretical discount.
