Anthropic has confirmed that Claude Sonnet 5 will not get more expensive on 1 September 2026. The model's launch pricing of $2 per million input tokens and $10 per million output tokens, originally framed as a limited-time introductory rate, is now the standard price. A previously scheduled rise to $3 input and $15 output, due to take effect the day this article is dated, will not happen.
This is a direct reversal of a plan that was public since Sonnet 5 launched, and it is worth understanding exactly what changed, what didn't, and what it means if you're running agent skills against the model through the Claude API, Claude Code, or Claude Managed Agents.
What Anthropic's pricing page actually says
Anthropic's own pricing reference carries a note attached directly to the Claude Sonnet 5 row of its model pricing table:
The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.
That is about as unambiguous as a pricing change gets: not a discount extension, not a delay, a cancellation of the increase entirely. The note is dated to sit right up against 31 August 2026, the last day the introductory rate was originally supposed to apply, which matches the timeline Anthropic set when Sonnet 5 first shipped.
The current pricing table
As of this pricing page, here is where Claude Sonnet 5 sits against the current model lineup, all figures per million tokens:
| Model | Input | Output | Batch input | Batch output |
|---|---|---|---|---|
| Claude Fable 5 | $10 | $50 | $5 | $25 |
| Claude Opus 5 | $5 | $25 | $2.50 | $12.50 |
| Claude Sonnet 5 | $2 | $10 | $1 | $5 |
| Claude Sonnet 4.6 | $3 | $15 | $1.50 | $7.50 |
| Claude Sonnet 4.5 | $3 | $15 | $1.50 | $7.50 |
| Claude Haiku 4.5 | $1 | $5 | $0.50 | $2.50 |
The comparison that matters most is the one against Sonnet 5's immediate predecessors. Claude Sonnet 4.6 and Claude Sonnet 4.5 both price at $3/$15, the exact rate Sonnet 5 was scheduled to move to on 1 September. With that increase cancelled, Sonnet 5 now sits roughly a third cheaper than the generation it replaced, rather than converging on the same price as the original plan intended.
Prompt caching multipliers apply on top of the same base rate: a 5-minute cache write costs 1.25x the base input price, a 1-hour cache write costs 2x, and a cache read (hit) costs 0.1x. None of those multipliers changed. What changed is only the base rate they multiply against.
Corroborating evidence, from an unexpected place
The pricing page isn't the only place this shows up. Claude Code's own changelog independently confirms the same fact, in a release that shipped before this article's publish date. Version 2.1.243, released 25 August 2026, includes this line:
Updated the
/modelpicker and the bundledclaude-apiskill to show Sonnet 5's $2/$10 per Mtok pricing as its standard list price rather than a limited-time promo.
That's a second, independent primary source, Claude Code's own release notes, confirming the same change from a different angle: not just that the pricing page says the increase won't happen, but that Anthropic's own tooling has already been updated to stop describing $2/$10 as a promotion. If you use the Claude API skill inside Claude Code, this is why its guidance on Sonnet 5 costs no longer carries a "this is temporary" caveat.
Why this matters if you're running agent skills
This site's whole audience runs agent skills against Claude in one form or another, whether that's Claude Code sessions, a Claude Managed Agents deployment, or direct Claude API calls from your own tooling. Sonnet 5 is Anthropic's balanced, mid-tier model, the one most cost-sensitive automated workloads default to rather than the pricier Opus tier. A scheduled 50% price increase on your default model is the kind of thing worth budgeting around well in advance; a cancelled one is worth confirming before you make any spending changes you don't need to make.
Concretely, this affects:
- Direct Claude API usage billed by the token, at the base $2/$10 rate shown above.
- Claude Code sessions authenticated with an API key, on usage credits after exceeding a plan's included usage, or running on a cloud provider like Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry, all of which bill by the token at published rates.
- Claude Managed Agents sessions, which bill on tokens plus session runtime, with token costs following the same model pricing table.
- Claude Platform on AWS, which converts the same per-model rates into Claude Consumption Units rather than charging a different price outright.
If you'd already built a cost model assuming Sonnet 5 would jump to $3/$15 on 1 September, that model is now wrong in your favour. If you'd budgeted for the increase and were planning to move heavy workloads to Haiku 4.5 or shrink context to compensate, it's worth revisiting whether that migration is still worth doing on cost grounds alone.
How this fits into Anthropic's wider pricing pattern this year
Sonnet 5 isn't the only place Anthropic has touched pricing mechanics through 2026. Data residency pricing introduced a 1.1x multiplier for US-only inference on Claude 4.6-and-later models, covered in detail separately, and Claude Managed Agents introduced an entirely new billing dimension, session runtime billed per hour, alongside the familiar per-token rates. Both of those additions moved in the direction of more granular, usage-shaped pricing rather than a flat per-model rate. The Sonnet 5 reversal moves in the opposite direction: instead of adding a new dimension or multiplier, it removes a previously announced increase to the simplest number in the whole system, the base per-token rate. Read together, the pattern is less "prices are only ever going up" and more "the mechanics are getting more granular while the headline rate on the mid-tier model has, for now, gone the other way."
What didn't change
It's worth being precise about the boundaries of this reversal, since it's easy to over-read a "prices aren't rising" announcement as "nothing about pricing changed."
- Every other model's pricing is unaffected. Opus 5, Fable 5 and Haiku 4.5 are not mentioned in the note and their listed rates are unchanged.
- Cache and batch multipliers are unchanged. The 1.25x/2x/0.1x cache multipliers and the roughly 50% batch discount apply the same way to the same $2/$10 base.
- The 1.1x data residency multiplier for US-only inference still applies to Sonnet 5 the same as it does to any Claude 4.6-or-later model, on top of whichever base rate is in effect.
- This is Anthropic's own first-party pricing. Amazon Bedrock and Google Cloud set independent regional pricing for Claude models on their platforms; check Bedrock's own pricing page or Google Cloud's if you're running Sonnet 5 through one of those rather than the Claude API or Claude Platform on AWS directly.
What a cancelled increase looks like at scale
The size of the reversal is easier to see in absolute terms than in the abstract $2 vs $3 comparison. Take a workload processing 500 million input tokens and 100 million output tokens a month, not an unusual volume for a busy agentic pipeline running many skill invocations a day.
| At $2/$10 (current, permanent) | At $3/$15 (the cancelled plan) | Difference | |
|---|---|---|---|
| Input cost | $1,000 | $1,500 | $500 |
| Output cost | $1,000 | $1,500 | $500 |
| Monthly total | $2,000 | $3,000 | $1,000, a 50% increase |
That's the number that would have shown up on a bill from 1 September onward under the original schedule. At this kind of volume, prompt caching and batch discounts change the absolute figures substantially, but the proportional gap stays the same: whatever caching and batching bring your effective rate down to, the cancelled increase would have applied that same 50% multiplier on top of it. For a team running Claude Sonnet 5 as the default model behind dozens of scheduled Claude Managed Agents deployments or a fleet of self-hosted Claude Code runners, that's real budget headroom to keep rather than lose.
Why this is worth treating as news, not just a footnote
Anthropic doesn't often reverse a previously announced, dated pricing change outright. Models more commonly ship at one price and stay there, or get replaced by a newer model at a different price point when the old one is deprecated. A specific, scheduled increase being cancelled two months ahead of taking effect, and confirmed on the actual day it was due to land, is unusual enough that it's worth flagging even to readers who don't track Anthropic's pricing page closely. It's also a useful data point on how Anthropic treats introductory pricing generally: the $2/$10 rate was explicitly framed as temporary at launch, and became permanent instead of expiring on schedule, which is the opposite of what "introductory" pricing usually turns out to mean.
Where to verify this yourself
Don't take a summary article, including this one, as the last word on a pricing question. Anthropic's pricing reference is the primary source this article draws from, and it's worth bookmarking directly if you're managing spend on any Claude-based workload: it's also where cache pricing, fast mode pricing, tool-use token overhead and Claude Managed Agents' session-runtime billing all live, none of which are things a general news summary reliably keeps current.
Troubleshooting: my cost dashboard still shows the old assumption
If a spend forecast, budget alert, or internal cost calculator still assumes $3/$15 for Sonnet 5 starting 1 September, that's a stale assumption baked into whatever tool built it, not a sign the pricing page is wrong. Anthropic's modelPricing managed setting in Claude Code lets an organisation pin contracted rates centrally rather than relying on each developer's dashboard staying current by hand; outside Claude Code, the fix is simply updating whatever spreadsheet or internal tool encoded the old schedule.
Where to go next
For the mechanics of how Claude Managed Agents bills tokens and session runtime together, see how budgets work across Claude Managed Agents sessions and deployments. For a related settings-driven cost control inside Claude Code, see Claude Code's promptCacheTtl setting explained. For the open-source skill that teaches Claude to write code against Anthropic's own APIs and keeps this kind of pricing detail current inside your own sessions, see Anthropic's Claude API skill explained. Browse the full catalogue at getclaudeskills.com/skills.
Verified 31 August 2026 directly against Anthropic's pricing reference at platform.claude.com/docs/en/about-claude/pricing, and cross-checked against Claude Code's own changelog entry for v2.1.243 (25 August 2026), which independently confirms the same change.
