VynarisEarly betaGet your API key

Blog · 2026-09-10 · Vynaris Team

Sonnet 5 just got 50% more expensive: $3/$15 is live, the $2/$10 rate lasted 18 days

Anthropic live pricing page confirms Sonnet 5 at $3/$15 per MTok from Sep 1. The $2/$10 introductory rate lasted 18 days. An 8k/2k call now costs $0.054, up 50% from $0.036. Sonnet 5 matches Sonnet 4.6. Cache reads, writes, and Batch all rose 50%.

Claude Sonnet 5 now costs $3 per million input tokens and $15 per million output tokens. The live Anthropic pricing page confirms the increase took effect September 1, 2026. A balanced 8k-input/2k-output call costs $0.054, up from $0.036. Every line item rose 50%. Prices verified 2026-09-10.

TL;DR

Verdict table

Line item                    Intro (through Aug 31)  Standard (from Sep 1)  Change
---------------------------  ----------------------  ---------------------  -------
Input per MTok               $2                      $3                     +50%
Output per MTok              $10                     $15                    +50%
Cache read per MTok          $0.20                   $0.30                  +50%
Cache write 5m per MTok      $2.50                   $3.75                  +50%
Cache write 1h per MTok      $4                      $6                     +50%
Batch input per MTok         $1                      $1.50                  +50%
Batch output per MTok        $5                      $7.50                  +50%
1,000 calls, 8k in + 2k out  $36                     $54                    +$18
1,000 calls, cache-heavy     $25.20                  $37.80                 +$12.60
1,000 calls, Batch API       $18                     $27                    +$9

The Anthropic pricing documentation now carries two rows for Sonnet 5. One says "through August 31, 2026" at $2/$10. The other says "starting September 1, 2026" at $3/$15. A sentence below the table states: "Introductory pricing of $2/$10 per million input/output tokens is in effect through August 31, 2026, after which the standard pricing of $3/$15 per million input/output tokens will take effect."

The $18 calculation

Our balanced workload contains 8,000 input tokens and 2,000 output tokens per call. Across 1,000 calls, that is 8 million input and 2 million output tokens.

At the introductory rate: 8M x $2/M + 2M x $10/M = $16 + $20 = $36. At the standard rate: 8M x $3/M + 2M x $15/M = $24 + $30 = $54. The difference is $18 per 1,000 calls, or $0.018 per call.

The calculation is deliberately flat. It needs no traffic estimate, benchmark score, or internal data. Change the token shape in the calculator if your workload carries a longer context window or heavier output.

Sonnet 5 per-1,000-calls cost: intro vs standard
Source: Anthropic pricing page, verified 2026-09-10. All line items +50%.

What happened to the "cancellation"

On August 14, we published an article reporting that Anthropic had canceled the scheduled September 1 increase. That article quoted the live page as saying the $2/$10 price "is now the standard price" and that the $3/$15 increase "will not occur."

The live page no longer says that. It now labels $2/$10 as "introductory" pricing with an end date of August 31, and lists $3/$15 as the standard rate from September 1. We cannot tell from the page alone whether Anthropic reversed a cancellation or whether the August 14 page state was temporary. The page does not explain the change. We report what it shows today.

The practical result: anyone who budgeted Sonnet 5 at $2/$10 based on the August 14 page state is now paying 50% more. The $2/$10 rate that our article called permanent lasted 18 days.

Sonnet 5 now matches Sonnet 4.6

The standard $3/$15 rate is identical to Sonnet 4.6. Cache reads match at $0.30. Cache writes match at $3.75 (5m) and $6 (1h). Batch rates match at $1.50/$7.50.

This eliminates the price gap that existed during the introductory period. For 18 days, Sonnet 5 was the cheapest Sonnet ever at $2/$10. That window is closed. A model routing policy that chose Sonnet 5 over Sonnet 4.6 on price alone has no remaining cost basis.

The Anthropic tokenizer still matters here. Sonnet 5 uses a newer tokenizer that produces roughly 30% more tokens for the same text. At $2/$10, that token inflation was partly offset by the lower sticker. At $3/$15, the token inflation stacks on top of the higher rate. The real bill grows faster than the 50% headline suggests if your prompts are long.

Cache and batch impact

Prompt caching users face the same 50% increase on every cache column. Cache reads moved from $0.20 to $0.30 per MTok. A cache-heavy workload with 6 million cache-read tokens, 2 million uncached input tokens, and 2 million output tokens per 1,000 tasks costs $37.80, up from $25.20.

The cache-read multiplier stays at 10% of input price. At $3 input, cache reads are $0.30. At $2, they were $0.20. The ratio is unchanged. What changed is the absolute dollar cost per cached token.

Our prompt caching break-even analysis and our cache-write churn cost study both used the $2/$10 rate. Their break-even thresholds shift proportionally. A cache write that needed 12.5 cache reads to break even at $2/$10 still needs 12.5 reads at $3/$15. The ratio is stable. The dollar amounts are 50% higher.

The Batch API discount remains 50% off standard rates. Batch input is $1.50 per MTok, output is $7.50. A balanced 1,000-call batch workload costs $27, up from $18. Batch is still the cheapest Sonnet 5 path when latency is acceptable.

Agent workloads take the largest dollar hit

A single agent step with 50,000 input and 15,000 output tokens now costs $0.375, up from $0.25. Across 1,000 steps, that is $375 instead of $250. The $125 gap is the largest absolute increase in our modeled workloads.

Our coding agent cost-per-task leaderboard showed Claude Code at $1.47 per task on Kimi K3 list pricing. That figure used Sonnet 5 at $2/$10. At $3/$15, every Claude Code task that touches Sonnet 5 costs 50% more in token fees. The leaderboard's relative rankings do not change, but the absolute dollar column shifts upward.

Secondary: GLM-5.3 Flash promo also expired

The same week brought a second price change. GLM-5.3 Flash doubled from promotional to list rates on September 9 at 16:00 UTC. Input moved from $0.075 to $0.15 per MTok. Output moved from $0.25 to $0.50. A balanced 1,000-call workload costs $2.20, up from $1.10.

Two models, two price increases, one week. The GLM-5.3 Flash increase is 100%. The Sonnet 5 increase is 50%. In absolute dollars, Sonnet 5's $18-per-1,000-calls increase dwarfs GLM-5.3 Flash's $1.10 jump. In percentage terms, GLM-5.3 Flash buyers took the larger hit.

Honest tradeoff: the increase makes alternatives more attractive, not less

A 50% price increase on Sonnet 5 does not automatically justify a migration. Haiku 4.5 at $1/$5 is still one-third the input cost and one-third the output cost of standard Sonnet 5. If Haiku passes your quality bar on a given workload, the price gap just widened from 2x to 3x.

The mistake is routing on price alone. A cheaper model that fails the task costs more in retries, reviewer time, and user trust than the 50% increase. The correct response is to re-run your per-request cost evaluation at the new $3/$15 rate and let quality-gated routing decide which calls still belong on Sonnet 5.

The opposite mistake is ignoring the increase. Budgets built on $2/$10 are now underfunded by 50%. Update the rate in every forecast, alert threshold, and cost-per-query dashboard before the next invoice arrives.

FAQ

Is Sonnet 5 still cheaper than Sonnet 4.6?

No. Both are $3/$15 per MTok from September 1. Cache reads, cache writes, and Batch rates are identical. There is no price reason to prefer one over the other.

Does the increase affect cached requests?

Yes. Cache reads rose from $0.20 to $0.30 per MTok. Cache writes (5m and 1h) rose 50% as well. The cache-read multiplier stays at 10% of input price, but the absolute dollar cost per cached token is higher.

Does the Batch API discount still apply?

Yes. Batch is 50% off standard rates. At $3/$15, Batch costs $1.50/$7.50 per MTok. A balanced 1,000-call batch workload costs $27, up from $18.

What about the August 14 article that said the increase was canceled?

That article reported what the live page showed on August 14. The page now shows $3/$15 from September 1 and labels $2/$10 as introductory. We do not know whether Anthropic reversed a cancellation or whether the August 14 state was temporary. The page does not explain the change.

Should I switch to a different model?

Re-run your quality-gated evaluation at $3/$15. Haiku 4.5 at $1/$5 is now 3x cheaper per token. Opus 5 at $5/$25 is 1.67x more expensive. The right answer depends on your task's quality bar, not on the sticker alone.

Sources