Claude Sonnet 5: What's New, What It Costs, and How to Switch Your Agents
Claude Sonnet 5 (model id claude-sonnet-5) launched June 30, 2026 as the most agentic Sonnet yet, performing close to Opus 4.8 at lower prices and now the default on Free and Pro. It runs $2/$10 per MTok intro through Aug 31, then $3/$15, with 1M context, 128k output, and adaptive thinking (effort defaults high).
What just launched (and why builders should care)
Claude Sonnet 5 is Anthropic’s new value-tier model, live since June 30, 2026 across every Claude app and the Claude Platform/API. The headline for builders: near-Opus agentic capability at Sonnet prices — $3 / $15 per MTok standard ($2 / $10 introductory through Aug 31, 2026), a 1M-token context window, 128k max output, adaptive thinking on by default, and it’s now the default model on Free and Pro.
Here’s why I’d clear an afternoon for it. Anthropic calls it “the best combination of speed and intelligence” and “the most agentic Sonnet yet”: it plans, drives tools like browsers and terminals, runs autonomously, and — the part that matters in production — checks its own output without being asked. Early-access partners report it finishes complex tasks where previous Sonnets quietly stopped short. It’s available today, no waitlist, and shipped as the default on Free and Pro, so most of your users are already talking to it whether or not you’ve touched a line of code.
My honest framing, not a press release: the real story is cost times agency. You can now run autonomous, tool-using, self-checking agents at Sonnet prices — autonomy that a few months ago meant paying Opus rates. For a small team in Puebla shipping production agents, that math changes what you can afford to build.
This post is the operational take: the facts you need to make a call, the five-step migration from Sonnet 4.6, and the decision frame for Sonnet 5 versus Opus 4.8 versus Haiku 4.5. I’ll be blunt about where Opus still wins and where the intro pricing has a clock on it.
The facts you actually need: id, price, specs
The model id is claude-sonnet-5. That single string is both the API id and the alias — it’s a dateless pinned-snapshot format, so do not append a date suffix. On Bedrock it’s anthropic.claude-sonnet-5; on Google Cloud it’s claude-sonnet-5.
Pricing, as of launch — confirm current rates before you commit: standard is $3 per input MTok and $15 per output MTok. There’s an introductory price of $2 / $10 through August 31, 2026, after which it reverts to $3 / $15. At standard pricing that’s roughly 60% cheaper per token than Opus 4.8, which runs $5 / $25.
Specs: a 1M-token context window; 128k max output on the synchronous Messages API (up to 300k on the Batch API with the output-300k-2026-03-24 beta header); a reliable knowledge cutoff of January 2026; and comparative latency rated “Fast”.
On the API surface, Sonnet 5 uses the modern adaptive-thinking surface — the same one as Opus 4.8. Adaptive thinking is on by default: a request with no thinking field already runs with adaptive reasoning, so you don’t have to enable it. You can set it explicitly with thinking: {"type": "adaptive"} for clarity, and pass thinking: {"type": "disabled"} to turn thinking off. Control depth with output_config.effort — the levels are low / medium / high / xhigh / max — which defaults to high on the Claude API and Claude Code. Note xhigh: it sits between high and max, and Anthropic recommends it for the hardest coding and agentic work. More on effort in the migration section, because the default is the thing that bites people moving from older models.
Sources to bookmark: Anthropic — Introducing Claude Sonnet 5, Models overview, and Pricing.
How agentic is it, really? The honest benchmark read
Let me give you the gap up front so you can trust the rest of this post. These are Anthropic’s reported figures, and the precise numbers live in the System Card, not the announcement page — so confirm the exact eval names and values there before you quote them. I’m describing the shape, not memorizing labels.
On an agentic-coding evaluation, Sonnet 5 lands in the low-60s percent, versus roughly the high-60s for Opus 4.8 and high-50s for Sonnet 4.6 (about 63.2% / 69.2% / 58.1% as reported). So Sonnet 5 jumps a meaningful step over the previous Sonnet while still trailing Opus on the hardest coding. Anthropic also evaluated it on BrowseComp (agentic search) and OSWorld-Verified (computer use) at different effort levels, where it approaches Opus across effort settings on search.
My read, labeled as opinion: Sonnet 5 closes most of the gap to Opus 4.8 at a fraction of the cost, but Opus still leads on the hardest agentic coding and long-horizon work. Be honest with yourself about positioning — Sonnet 5 is the value-tier sweet spot, not the absolute frontier. That crown belongs to Opus 4.8 and Fable 5. Confirm the eval names yourself before you quote a number to a stakeholder.
Why this matters for a builder here: you can ship autonomous, tool-using, self-checking agents without paying Opus rates. For most of the agent work I do — research loops, browser automation, code edits with a verify step — Sonnet 5 is now the obvious default, and I escalate only the genuinely hard tasks. More patterns in my AI agents guides.
Switch your agents in five steps
The migration is small. Here’s the whole thing in five moves.
Step 1 — change the model string to claude-sonnet-5. A minimal Messages API call looks like this:
import anthropic
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=2048,
thinking={"type": "adaptive"}, # optional — adaptive is on by default
output_config={"effort": "high"}, # defaults to high; set explicitly to change
messages=[
{"role": "user", "content": "Plan and run the steps to triage this bug report."}
],
)
print(message.content)
print(message.usage) # input_tokens, output_tokens — track your real spendStep 2 — use adaptive thinking. Extended thinking with budget_tokens is not supported on this surface — passing it returns a 400 — and Sonnet 5 is adaptive-only, on by default. If you were passing budget_tokens, drop it; the explicit thinking: {"type": "adaptive"} line is optional but harmless for clarity. To turn thinking off entirely, pass thinking: {"type": "disabled"}. Here’s the diff from Sonnet 4.6:
message = client.messages.create(
- model="claude-sonnet-4-6",
+ model="claude-sonnet-5",
max_tokens=2048,
- thinking={"type": "enabled", "budget_tokens": 8000},
+ thinking={"type": "adaptive"},
+ output_config={"effort": "high"},
messages=messages,
)Step 3 — re-tune effort. output_config.effort defaults to high on the Claude API and Claude Code. Set it explicitly — the levels are low / medium / high / xhigh / max — to change. Re-tune deliberately, because the default shifted versus older models: if your agent suddenly thinks harder and costs more per call than you expected, this is usually why. For the hardest coding and agentic tasks, Anthropic recommends stepping up to xhigh.
Step 4 — drop unused sampling params. temperature, top_p, and top_k are not used on this surface. Remove them rather than leaving dead config around.
Step 5 — pilot during the intro pricing window (through Aug 31) and add prompt caching to cut the bill. Cache the stable parts of your prompt — system instructions and tool definitions — with cache_control:
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=2048,
thinking={"type": "adaptive"},
system=[
{
"type": "text",
"text": LONG_SYSTEM_PROMPT_AND_TOOL_CONTEXT,
"cache_control": {"type": "ephemeral"},
}
],
messages=messages,
)Everything else is the standard Messages API — tool use and structured outputs via output_config.format. Confirm the exact breaking changes in Anthropic’s migration guide before you ship to production.
- Change the model stringSet model to claude-sonnet-5 (no date suffix)
- Use adaptive thinkingAdaptive is on by default; drop budget_tokens
- Re-tune effortDefaults high — set low/medium/high/xhigh/max explicitly
- Pilot in the intro windowRun real workloads while it's $2 / $10
- Add prompt cachingcache_control on system + tools to cut the bill
Sonnet 5 vs Opus 4.8 vs Haiku 4.5: which do I pick?
Route by task, not by habit. Here’s how I split the three.
Pick Sonnet 5 for agentic workloads at scale — autonomous, tool-using, self-checking agents at Sonnet prices. This is the value-times-agency sweet spot and where most of my production traffic now lands.
Pick Opus 4.8 ($5 / $25) when you need the absolute top: the hardest agentic coding, long-horizon work where a single wrong step compounds. It still leads, and on a critical task the extra few points are worth the rate.
Pick Haiku 4.5 ($1 / $5) for the cheapest, fastest simple calls — classification, extraction, short rewrites. Note the tradeoffs: 200k context, 64k output, a February 2025 knowledge cutoff, and it uses extended thinking rather than the adaptive surface, so it’s not a drop-in for adaptive code.
My rule of thumb: default to Sonnet 5, escalate specific hard tasks to Opus, downshift trivial calls to Haiku. If you’re weighing the absolute frontier, I broke down how the top models compare separately.
Claude Sonnet 5
- Near-Opus agency and coding at ~60% lower cost
- Default on Free and Pro
- Adaptive thinking on by default, effort defaults high
- Best for autonomous, tool-using agents at scale
Claude Opus 4.8
- Leads on the hardest agentic coding
- Best for long-horizon, high-stakes work
- $5 / $25 per MTok — the premium tier
- Reach for it when a wrong step compounds
The cost play: pilot now, cache aggressively
The intro window — $2 / $10 — runs through August 31, 2026, then reverts to $3 / $15. So pilot real workloads now, while the per-token math is cheapest. But build your unit economics assuming standard pricing. Model the bill at $3 / $15 so nothing breaks on September 1, and treat the intro discount as a bonus, not a foundation.
The single biggest lever in an agent loop is prompt caching. Agent calls re-send the same system prompt and tool definitions on every turn — cache them with cache_control (as in Step 5 above) and you stop paying full input price for context that never changes. On a chatty agent, that’s the difference between a sustainable bill and a scary one. I wrote a full walkthrough on cutting token costs with prompt caching.
For batch or offline jobs, the Batch API unlocks up to 300k output tokens with the output-300k-2026-03-24 beta header, plus batch discounts — worth it for anything that doesn’t need a synchronous response.
The practical LATAM angle: cost times agency means small teams can now run production agents that needed Opus-tier budgets a few months ago. That’s not hype — it’s the difference between “we’ll prototype this someday” and “this is live this week.” For the broader picture, here’s my running guide to running the model in production.
FAQ
What’s the exact model id? claude-sonnet-5 — no date suffix. It’s a dateless pinned snapshot, and the same string works as both the API id and the alias.
Does it support budget_tokens or extended thinking? No. Sonnet 5 is adaptive-thinking only, and adaptive is on by default. Passing budget_tokens returns a 400 — control depth with effort instead, and pass thinking: {"type": "disabled"} to turn thinking off.
What does effort default to, and what are the levels? It defaults to high on the Claude API and Claude Code. The full set is low / medium / high / xhigh / max — set output_config.effort explicitly to use anything other than high, and reach for xhigh on the hardest coding and agentic tasks.
Is the $2 / $10 price permanent? No — it’s introductory through August 31, 2026, then reverts to $3 / $15. Confirm current rates at the Pricing docs.
How big is the context and output? 1M-token context, 128k max output on the Messages API (300k on the Batch API with the beta header), with a January 2026 knowledge cutoff.
Can I set temperature, top_p, or top_k? No — those sampling params are not used on this surface.
Bottom line
Sonnet 5 is the value-times-agency sweet spot: near-Opus autonomy at roughly 60% lower cost, and it’s already the default on Free and Pro. The migration is small — change the model string to claude-sonnet-5, lean on adaptive thinking (it’s on by default), drop budget_tokens, and re-tune effort across low / medium / high / xhigh / max since it now defaults to high. Pilot during the intro window, cache aggressively, and model your bill at standard pricing so the numbers survive September 1. Reach for Opus 4.8 only when you genuinely need the frontier, and keep Haiku 4.5 around for the cheap, simple calls. For most agent work, Sonnet 5 is now my default — and I think it’ll be yours too.