The most useful thing about Claude Sonnet 5 is not that it plans multi-step tasks better than its predecessor. It is that on August 10, 2026, Anthropic made the cheap pricing permanent.
That sounds like a boring accounting footnote. For anyone who actually ships agents, it is the detail that changes architecture.
What actually shipped
Sonnet 5 landed on June 30, 2026. Anthropic called it its most agentic Sonnet model to date, built to plan multi-step work and operate tools with less human babysitting. On July 1 it became the default model for Free and Pro users, pushing Sonnet 4.6 out of the driver’s seat. It carries a 1M-token context window.
Pricing came in at $2 per million input tokens and $10 per million output tokens, positioned below Anthropic’s flagship Opus model through August 31. The published plan was that the rate would step up to $3 and $15 after the promo window closed. That step-up never happened. On August 10, the introductory numbers became the permanent ones.
So the story most people are retelling — “cheap launch pricing, enjoy it while it lasts” — is already out of date.
Why permanence is the feature
When I design a bot, the token price is not a line item I check once. It is a constraint baked into every loop I write. A promotional rate with an expiry date forces defensive design. You build the agent that survives the price hike, not the agent that does the best job. That usually means:
- Fewer retries, because each retry is a bet on a rate you might not have in six weeks
- Aggressive context trimming, even when the extra context would improve the output
- Skipping verification passes, because a second model call to check the first one doubles your exposure
- Routing more work to smaller models than the task really deserves
Every one of those choices makes the agent worse. They are hedges against a pricing decision, not engineering decisions.
Make the rate permanent and the hedges go away. I can write a plan-execute-verify loop where verification is a real step instead of an aspiration. I can hand an agent the full repository context instead of a clever summary of it that loses the one function it needed. I can let a tool-use loop run to eight or ten turns without watching the meter.
Agentic and cheap pull in the same direction
Here is the part that I think gets underrated. Agentic capability and low token cost are not two separate improvements you evaluate on separate axes. They compound, and they compound badly if only one of them shows up.
An agentic model is, by definition, a model that burns more tokens per unit of work. It plans, then it calls a tool, then it reads the result, then it revises the plan. A single user request can turn into a dozen model calls. That is the whole point — the reduced supervision Anthropic describes comes from the model doing the iterating instead of you doing it.
Which means an expensive agentic model is a demo, not a product. The capability exists but the unit economics stop you from letting it run. You end up capping the loop at three turns and calling it an agent, when what you have built is a slightly chatty function call.
Pairing the agentic push with a price that stays put is what makes the loop worth writing. The 1M-token context window sits in the same bucket. A million tokens of context you cannot afford to fill is a spec sheet number. A million tokens at $2 per million input is a design option.
The default-model switch matters too
Rolling Sonnet 5 out as the default for Free and Pro on July 1 is a distribution decision, but it has a practical consequence for builders. When a mid-tier model becomes what most people are talking to, the behavior of that model becomes the baseline expectation your users bring to your bot. If your agent is running something older, the gap shows up as your product feeling slow-witted, not as the model feeling old.
What I would change in my own stack
Concretely, three things I am revisiting:
- Retry and self-check budgets. If a task is worth doing, it is worth a verification pass. I had these tuned down for cost reasons that no longer apply.
- Model routing rules. Some of my routing logic exists purely to keep expensive calls rare. With Sonnet 5 sitting below Opus on price and carrying stronger planning behavior, a chunk of that routing is just complexity now.
- Context strategy. I have a lot of summarization machinery that exists to squeeze inputs. Some of it is still valuable for latency and relevance. Some of it was pure cost engineering and can go.
None of that is glamorous. It is mostly deleting workarounds. But that is usually what a genuinely useful model release looks like from inside a codebase — not new capabilities you rush to add, but old compromises you finally get to remove.
Check the rates yourself before you commit a budget. Pricing pages move, and the interesting thing about this one is that it moved in the direction nobody expected.
🕒 Published: