LLM cost planning autumn 2026: Sonnet 5 stays at $2/$10

Anthropic confirmed on 10 August 2026 that Claude Sonnet 5's introductory pricing of $2 per million input tokens and $10 per million output tokens is now the standard price. The increase to $3/$15 per million tokens that had been scheduled for 1 September 2026 will not happen. Anthropic's pricing documentation states it plainly: "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."

For teams running on Azure this is directly relevant, because Claude stopped being an off-Azure exception this summer. Claude in Microsoft Foundry went generally available on 29 June 2026, and Claude Sonnet 5 followed into Foundry at launch on 30 June, with Azure-native authentication through Entra ID, Azure role-based access control, and billing that lands on your Azure invoice. Any Swedish or EU enterprise that budgeted a 50% Sonnet 5 price rise into its September forecasts can now take that line back out, and the change also shifts the mid-tier comparison against GPT-5.6 and Gemini. The numbers below reflect published list prices as of 12 August 2026.

What was scheduled, and what got cancelled

Claude Sonnet 5 launched on 30 June 2026, positioned by Anthropic as "the most agentic Sonnet model yet," with improvements over Sonnet 4.6 in reasoning, tool use, coding and knowledge work. At launch, Anthropic set introductory pricing of $2/$10 per million tokens, valid through 31 August 2026, after which the price would move to $3/$15: the same list price as Sonnet 4.6 and Sonnet 4.5 before it. Microsoft's own Foundry announcement repeated the schedule, so plenty of Azure teams wrote the September increase into their cost models.

The 10 August announcement cancels that schedule. Sonnet 5 stays at $2/$10 indefinitely. Derived prices move with it: batch processing keeps its 50% discount at $1/$5 per million tokens, prompt cache reads stay at $0.20 per million tokens (10% of base input), and 5-minute cache writes stay at $2.50 per million tokens. Sonnet 4.6 remains available at $3/$15, which now makes the older model strictly more expensive than its successor on list price.

The cost math: what the cancelled increase was worth

Take a realistic production workload: an internal assistant or document pipeline consuming 500 million input tokens and 50 million output tokens per month on Sonnet 5.

At the confirmed $2/$10 price: 500 × $2 + 50 × $10 = $1,500 per month, roughly $18,000 per year.
At the cancelled $3/$15 price: 500 × $3 + 50 × $15 = $2,250 per month, roughly $27,000 per year.
The cancellation is worth about $9,000 per year per workload of this size, a 33% reduction against what many teams had budgeted from September onward.

The levers on top of base price are unchanged and still stack. Batch processing halves both input and output cost for anything that tolerates asynchronous turnaround, which covers most document processing, classification and evaluation workloads. Prompt caching cuts repeated context to 10% of the input price on cache hits, and a 5-minute cache write pays for itself after a single hit. The same 500M/50M workload run through the Batch API costs $750 per month. If 70% of your input tokens are cache reads in an interactive agent, the input side drops from $1,000 to around $370 before write costs. None of this is new, but every one of these percentages now applies to a smaller base than your September forecast assumed.

Price/performance across the tiers, August 2026

The decision confirms a pattern across vendors this summer: the capable mid-tier has converged on $2 per million input tokens. OpenAI launched the GPT-5.6 family on 9 July 2026 and cut prices on 30 July, dropping GPT-5.6 Terra by 20% to $2/$12 and GPT-5.6 Luna by 80% to $0.20/$1.20. Google lists Gemini 3.1 Pro at $2/$12 for prompts up to 200k tokens ($4/$18 above that), and Gemini 3.6 Flash at $0.75/$3.75, rates its pricing page marks as promotional through 31 December 2026 with a doubling to $1.50/$7.50 from January 2027. List prices per million tokens, as of 12 August 2026:

Model	Input / MTok	Output / MTok	Pricing note
Claude Opus 4.8	$5.00	$25.00	Standard
GPT-5.6 Sol	$5.00	$30.00	Standard
Claude Sonnet 5	$2.00	$10.00	Confirmed permanent 10 Aug
GPT-5.6 Terra	$2.00	$12.00	After 30 July cut
Gemini 3.1 Pro	$2.00	$12.00	≤200k context; $4/$18 above
Claude Sonnet 4.6	$3.00	$15.00	Now above its successor
Claude Haiku 4.5	$1.00	$5.00	Standard
Gemini 3.6 Flash	$0.75	$3.75	Promotional through Dec 2026
GPT-5.6 Luna	$0.20	$1.20	After 30 July cut

At the $2 input tier, Sonnet 5 now has the lowest output price of the three. Output tokens dominate cost in generation-heavy workloads such as code generation, drafting and long-form summarisation, so the $10 versus $12 gap matters more than it looks. For input-heavy retrieval workloads, where you push large contexts in and get short answers out, the three are effectively at parity on list price and your choice should rest on quality evaluations, caching behaviour and platform fit rather than the rate card.

The tokenizer caveat: per-token price is not per-task cost

One correction to apply before you paste these numbers into a spreadsheet. Anthropic documents that Claude models from 4.7 onward, including Sonnet 5, use a newer tokenizer that produces roughly 30% more tokens for the same text than the tokenizer used by Sonnet 4.6 and earlier. That means the move from Sonnet 4.6 at $3/$15 to Sonnet 5 at $2/$10 is not a clean 33% cost cut for identical workloads; the effective saving on the input side is smaller once the higher token count is applied. The same warning holds across vendors: OpenAI, Google and Anthropic tokenize differently, so a dollars-per-million-token table understates or overstates real cost depending on your content. The only reliable comparison is dollars per completed task, measured by running your own representative workload through each candidate model and reading actual token counts from the usage fields in the responses.

Running Sonnet 5 on Azure: how the price flows through Foundry

Since the 29 June general availability of Claude in Microsoft Foundry, Claude models are a production option inside the Azure ecosystem rather than a separate vendor relationship. Authentication runs through Microsoft Entra ID, access is governed by Azure RBAC and existing policy, and usage appears in Azure Cost Management. For high-sensitivity workloads, Microsoft's announcement notes a zero-data-retention option under which prompts and completions are not retained by Anthropic after the API call completes.

Billing works through the Azure Marketplace in Claude Consumption Units (CCUs). Anthropic rates your token usage in USD at the standard per-model prices, applies any negotiated discount, and converts the result at $0.01 per CCU; 100 CCU represents $1.00 of usage. The practical consequence of the 10 August decision is that the rate used in that conversion stays at $2/$10 for Sonnet 5. No contract change, no SKU swap, no action on your side: the September increase simply never reaches your Azure invoice. If you maintain internal chargeback models that translate CCU line items back to per-team token budgets, update the assumed Sonnet 5 rate now so September forecasts are not silently 50% too high.

One planning note for teams comparing hosting routes: Claude is also available through Amazon Bedrock and Google Vertex AI, where regional endpoints (the kind you choose when you need guaranteed geographic routing) carry a 10% premium over global endpoints for Claude 4.5-generation models and later. If EU data routing is a requirement, price that premium in from the start rather than discovering it at procurement review.

A decision framework for autumn model budgeting

Four questions to run against every LLM line item in your 2027 budget draft:

1. Is the quoted price introductory? Sonnet 5's introductory price became permanent, but that was Anthropic's choice, not a rule. Gemini 3.6 Flash's current rate is published as doubling in January 2027. Budget at the documented standard rate and treat promotional pricing as upside.
2. Is the workload output-heavy or input-heavy? Output-heavy workloads should weight the output rate, where Sonnet 5's $10 leads its tier. Input-heavy workloads should weight caching economics, since cache reads at 10% of input price dwarf any list-price difference between vendors.
3. Does the task need the top tier at all? The gap between Opus 4.8 at $5/$25 and Sonnet 5 at $2/$10 is 2.5x. Route by task: reserve the top tier for the hardest reasoning steps and let the mid-tier carry volume. A router prompt plus an evaluation set settles this empirically in a week.
4. Can it run in batch? Anything without a human waiting on the response qualifies for the 50% batch discount. Nightly document processing on Sonnet 5 at $1/$5 undercuts most dedicated cheap-tier models on quality-adjusted cost.

Rerun the arithmetic whenever a vendor moves. This summer alone produced a launch (GPT-5.6 on 9 July), a price cut (30 July) and a cancelled increase (10 August). A cost model reviewed quarterly is stale; monthly is the right cadence while the market moves at this pace.

The Swedish and EU angle

Budget season timing is favourable. Most Swedish enterprises and public-sector bodies build 2027 budgets between late August and October. The cancellation landed just before that window, so AI platform lines can be set against the confirmed $2/$10 rate instead of the $3/$15 assumption. For organisations invoicing through the Azure Marketplace, note also that CCU conversion is dollar-denominated; SEK exposure sits in the USD exchange rate, and that risk belongs in the same budget note as the token forecast.

Procurement should treat introductory pricing as a named risk. Upphandling processes routinely lock evaluation on quoted unit prices. This summer shows how unstable those quotes are: one vendor cancelled an increase, another cut prices three weeks after launch, a third publishes a doubling with a fixed date. When you evaluate bids that pass LLM costs through, require the documented standard rate, the promotional rate and its expiry date as separate fields, and score on the standard rate. Suppliers quoting only the promotional number are shifting repricing risk onto you.

Compliance overhead is now part of unit cost. Since 2 August 2026, the bulk of the EU AI Act's obligations apply, including requirements for deployers of high-risk systems. For affected use cases, evaluation runs, logging and human-oversight processes consume real tokens and real engineering time. A cost model that prices only production inference understates the true figure; evaluation and monitoring runs consume tokens in proportion to how often you ship changes, and that share should be budgeted as its own line rather than absorbed. Cheaper Sonnet 5 tokens make systematic evaluation cheaper too, which removes one excuse for skipping it.

Data residency options differ by route. Foundry offers a US Data Zone deployment type, and the first-party API routes globally by default. Teams with strict EU routing requirements currently get the most explicit geographic guarantees for Claude through Bedrock or Vertex regional endpoints at the 10% premium noted above. Document which route each workload uses and why; that mapping is exactly what a records-of-processing review or a security audit will ask for.

Takeaways

Remove the September increase from your forecasts. Sonnet 5 stays at $2/$10 per million tokens; a 500M/50M monthly workload stays at about $1,500 instead of $2,250.
Update chargeback and unit-economics models now, before 2027 budget drafts freeze the wrong rate.
Compare on dollars per task, not dollars per token. The 4.7+ tokenizer produces roughly 30% more tokens for the same text, and cross-vendor tokenizers differ further.
Weight output price for generation-heavy work. Sonnet 5's $10 output rate leads the $2-input tier against GPT-5.6 Terra and Gemini 3.1 Pro at $12.
Score procurement bids on standard rates, with promotional prices and expiry dates disclosed separately.
Exploit batch and caching before switching models. A 50% batch discount and 90% cache-read discount usually outweigh any list-price gap between comparable mid-tier models.
Budget AI Act evaluation and logging as an explicit line, not as a rounding error inside production inference.