Sonnet 5 Price Freeze [What It Changes]

TL;DR

• The Sonnet 5 Price Freeze is Anthropic locking Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, and cancelling the $3/$15 jump that had been set for September 1, 2026.

• As of August 10, 2026, that introductory rate is permanent on the Claude API; the September 1 “standard” price no longer applies.

• The freeze does not undercut OpenAI’s cheapest flagship or DeepSeek on raw list price; it removes a hike shops had already been told to plan for.

• Batch discounts, prompt-cache hits, and effort settings still move the bill more than the freeze itself, so the lock is a ceiling, not a finished cost model.

Key Takeaways

• If you already pay a vendor or developer to run tools on Claude Sonnet 5, the line item that was scheduled to jump 50% on September 1 stays at the June launch rate.

• API list price, consumer chat plans, and cloud-marketplace invoices are different bills; do not copy the $2/$10 pair onto a Pro seat or a Bedrock invoice without checking that invoice.

• Cache reads, batch jobs, and lower effort cut tokens (and dollars) without waiting for another vendor announcement.

• The freeze is a pricing decision, not a new model. Safety defaults, tokenizer behavior, and rate limits did not reset on August 10.

• Small shops should treat this as permission to keep the current model ID in production and spend the saved 50% on logging, evals, and a fallback route.

What the Sonnet 5 Price Freeze is

The Sonnet 5 Price Freeze is Anthropic’s decision, posted as an August 10, 2026 edit on the Claude Sonnet 5 launch note, to keep the model at $2 per million input tokens and $10 per million output tokens and to drop the $3 input / $15 output “standard” rate that had been scheduled to take effect on September 1.

A 2-truck HVAC shop that pays a developer to turn after-hours voicemails into dispatch notes does not buy GPUs. It buys a monthly software line. If that line was about to jump 50% because the model underneath was scheduled to reprice, the owner either raises service fees, cuts night coverage, or eats the cost. A 10-person marketing agency that drafts client recaps and CRM notes on Sonnet 5 had the same problem: a September hike would have shown up as a vendor invoice, not as a research paper. A solo-run clinic that summarizes intake forms into the chart has even less slack; one model bill is often the whole AI budget.

That is why the freeze matters before any benchmark chart does. It is a cost lock on a model already sitting inside production tools, not a new capability drop. Teams that already map small-business automation onto a single model ID can leave that ID in place instead of ripping out a stack to dodge a September invoice.

According to Anthropic’s Claude Sonnet 5 announcement, the model is priced at $2 per million input tokens and $10 per million output tokens as of June 2026, and $2/$10 per million tokens is now permanent. Third-party recap pages such as explainx.ai’s Sonnet 5 pricing note circulated around the same news cycle; treat Anthropic’s own changelog as the source of record.

The 42-day window

Claude Sonnet 5 shipped on June 30, 2026. The launch post presented cost-performance charts at $3 per million input tokens and $15 per million output tokens, labeled as “existing standard pricing,” while the live API rate was the introductory $2/$10 pair. The same post later records the freeze: the $3/$15 schedule set for September 1 no longer applies.

The cancelled hike was $3 / $15, as written in that August 10, 2026 changelog. Forty-two days separated launch and the lock. Shops that had already keyed $3/$15 into a fall forecast can now reverse that line. Shops that never saw the footnote still get the cheaper rate, but they should confirm the model ID claude-sonnet-5 on the invoice so a vendor is not billing an older Sonnet SKU.

The Sonnet 5 model page repeats the locked pair ($2 / $10), lists a 1M-token context window, 128K max output, a January 2026 knowledge cutoff, a June 30, 2026 release, and a retirement commitment of not sooner than June 30, 2027. That retirement floor is the closest thing to a “how long does this SKU last” answer Anthropic publishes; it is not a price guarantee beyond the freeze language.

Sources: Anthropic Claude Sonnet 5 launch and August 10 edit; Sonnet 5 model overview.

Predecessor pricing is the reason the cancelled $3/$15 pair looked familiar. Introducing Claude Sonnet 4.6 kept Sonnet 4.6 at $3/$15 per million tokens, matching Sonnet 4.5. The freeze is Sonnet 5 staying one step below that family, not a rewrite of the older SKUs.

How the locked rate is actually billed

List price is the starting line, not the invoice. The Claude API pricing page prices Claude Sonnet 5 at $2 / MTok input, $2.50 / MTok for 5-minute cache writes, $4 / MTok for 1-hour cache writes, $0.20 / MTok for cache hits, and $10 / MTok output. The same table still lists Claude Sonnet 4.6 at $3 / $15 and Claude Haiku 4.5 at $1 / $5, so a “Claude” invoice can hide three different rates.

Prompt caching is the lever most small tools should turn first. A 5-minute cache write is 1.25× base input; a cache hit is 0.1× base input on Sonnet 5 ($0.20). If a clinic sends the same intake template on every request, the template should be the cached prefix. If an agency sends a new brief every time and caches the changing block, it pays write price forever and never sees the $0.20 hit. The docs also note a 20-block lookback window: put the breakpoint on the last stable block, not on the timestamp.

Batch processing cuts standard API prices 50%. According to Anthropic’s API pricing table, Claude Sonnet 5 batch input is $1 / MTok and batch output is $5 / MTok. Batch processing cuts those rates 50%, which is the same math as $2×0.5 and $10×0.5. A batch is capped at 100,000 requests or 256 MB, most jobs finish in less than 1 hour, results stay downloadable for 29 days, and streaming, Fast mode, and max_tokens: 0 are not allowed inside a batch. Overnight CRM backfills and document-classification dumps belong here; live dispatch does not.

Effort is the other hidden multiplier. Sonnet 5 defaults to high on the Claude API and Claude Code. medium is the documented cost step-down; low is for high-volume or latency-sensitive chat; xhigh and max spend more tokens on purpose. Effort applies to text, tool calls, and thinking. A shop that leaves the default at high for every “summarize this voicemail” job will not see the freeze’s full benefit, because the freeze locks the rate, not the token count.

Context windows on Sonnet 5 are 1M tokens at standard pricing; you do not need a beta header, and a 900k-token request is billed at the same per-token rate as a 9k-token request. Cached tokens still occupy the window. Consumer Claude plans show a 200k context window on Free, Pro, Max 5x, and Max 20x seats, so do not assume the API’s 1M window is what a $20 Pro seat is using.

US-only inference is a priced add-on on the API: the pricing page applies a 1.1× multiplier on all token categories for Claude 4.6 and later when inference_geo is "us". Global routing is the default. Partner clouds (Amazon Bedrock, Google Cloud) publish their own regional rates; Anthropic points to Amazon Bedrock pricing and Google Cloud generative AI pricing rather than restating those invoices.

Sources: Anthropic API pricing; prompt caching; batch processing. The 1.1× row is $2.00×1.1 and $10.00×1.1 from the same pricing page.

How $2/$10 stacks against other APIs

The freeze is a lock, not a race win. According to OpenAI’s API pricing, gpt-5.6-luna short-context input is $0.20 per 1M tokens and output is $1.20. That is an order of magnitude below Sonnet 5’s frozen $2/$10. The same OpenAI table lists gpt-5.6-terra at $2.00 / $12.00 short-context, which is the closer mid-tier comparison, and gpt-5.6-sol at $4.00 / $20.00. Regional processing on eligible OpenAI models released on or after March 5, 2026 carries a 10% uplift on that pricing page.

According to DeepSeek’s Models & Pricing, deepseek-v4-flash off-peak cache-miss input is $0.22 per 1M tokens and off-peak output is $0.66, with peak hours 01:00–04:00 and 06:00–10:00 UTC Monday through Friday. Off-peak rates are half of peak. DeepSeek also publishes a 1M context length and a 384K max output. Raw dollars still favor DeepSeek over frozen Sonnet 5.

According to Google’s Gemini Developer API pricing, Gemini 3.8 Flash paid input is $0.75 per 1M tokens through December 31, 2026, then $1.50 starting January 1, 2027, with output $3.75 then $7.50. Google also states a Batch API 50% cost reduction on that page. Vertex / Agent Platform pricing repeats Gemini 3.7 Flash introductory $0.75 / $3.75 through December 31, 2026 and standard $1.50 / $7.50 from January 1, 2027.

Sources: Anthropic model overview; Anthropic API pricing; OpenAI API pricing; DeepSeek Models & Pricing; Gemini Developer API pricing. Em dashes mark context sizes not stated on the cited pricing tables.

The honest read for a shop already on Sonnet 5: you did not just get the cheapest model on the market. You got a cancelled 50% increase on a model your tools already call. Switching to luna, Flash, or DeepSeek is a product decision (quality, tools, data handling), not a freeze decision.

OpenAI’s data-controls guide states that as of March 1, 2023, data sent to the OpenAI API is not used to train models unless you opt in, with default abuse-monitoring retention of up to 30 days. That is a procurement checkbox, not a price. Anthropic’s consumer plans page lists model training as opt-out on Free through Max. Keep those rows next to the freeze when a clinic asks “is cheaper also safer.”

USTA analysis: the cancelled 50% delta

USTA analysis. This block uses only the locked $2/$10 and cancelled $3/$15 pair from Anthropic’s launch changelog. No other statistic is invented. A token mix of 5 input tokens per 1 output token is an illustration so a reader can check the arithmetic; replace the volumes with your own invoice.

Formula: monthly bill = (input millions × input $/MTok) + (output millions × output $/MTok). Avoided spend = cancelled-hike bill − locked bill. On this mix, input and output each move 50%, so the blended bill also moves 50%.

Worked row: 10 million input tokens and 2 million output tokens. Locked: (10 × 2) + (2 × 10) = $20 + $20 = $40. Cancelled hike: (10 × 3) + (2 × 15) = $30 + $30 = $60. Avoided: $20, which is 50% of $40 and 33% of $60.

Sources: rates from Anthropic Claude Sonnet 5; batch row uses the 50% Message Batches discount on API pricing ($1/$5 vs $1.50/$7.50). Volumes are illustrative; plug in your logs.

The last row is the same 10/2 mix on the Batches API: locked batch $1/$5 → $10 + $10 = $20; the cancelled standard $3/$15 would have batched to $1.50/$7.50 → $15 + $15 = $30. Overnight jobs that can wait an hour capture that extra 50% on top of the freeze.

One caveat sits in the launch footnotes: Sonnet 5 uses an updated tokenizer, and the same input can map to roughly 1.0–1.35× as many tokens depending on content type, per the launch post. The freeze locks dollars per token, not tokens per page. If a clinic’s “10 million tokens” was estimated on Sonnet 4.6, the Sonnet 5 bill can still rise even at $2/$10. Count tokens on the new tokenizer before you call the freeze a cut.

What the freeze does not change

Safety defaults did not thaw. The Claude Sonnet 5 System Card (June 30, 2026) is the capability and safety record. The launch post states that on a Firefox 147 exploit evaluation, neither Sonnet 5 nor Sonnet 4.6 developed a full working exploit (both 0.0%), that cyber safeguards launched enabled by default, and that those safeguards match Opus 4.7 / 4.8 rather than the stricter Fable 5 set. Anthropic’s help center documents the real-time cyber safeguards program; organizations already in it keep access on Sonnet 5 without reapplying.

Agentic evals that justified the model in the first place also stay. Anthropic compared Sonnet 5 to Sonnet 4.6 and Opus 4.8 on BrowseComp (1,266 questions for browsing agents, submitted 16 Apr 2025) and on OSWorld-Verified. The original OSWorld benchmark lists 369 computer tasks; the paper’s human baseline is over 72.36% success versus 12.24% for the best model in that original write-up. Anthropic later restated a Sonnet 4.6 OSWorld-Verified score of 78.5% after a methodology change, in the Sonnet 5 launch footnotes. Do not mix those two OSWorld generations in one cell.

Rate limits did not freeze. The rate-limits docs put Claude Sonnet 5 on its own bucket at 1,000 RPM, 2,000,000 ITPM, and 400,000 OTPM, separate from the combined Sonnet 4.x bucket. Monthly spend caps by tier are $500 (Start), $1,000 (Build), and $200,000 (Scale). Cached read tokens do not count toward ITPM on most Claude models. Hitting the cap returns HTTP 429 until 00:00 UTC on the first of the next month unless you raise the tier.

Consumer seats are a different SKU. On Claude plans, Pro is $17 per month when billed annually ($200 up front) or $20 monthly; Max starts at $100 per month; Enterprise is $20 per seat per month plus usage billed at API rates. The freeze language in the launch changelog is about API token rates, not about those seat prices. A shop on Pro that never touches the API did not just get a $2/$10 lock.

Sampling knobs tightened. The Sonnet 5 overview says adaptive thinking is on by default, manual extended thinking now returns 400, and non-default temperature / top_p / top_k return 400. A vendor who still sends temperature: 0.2 will fail after a “simple model swap” even though the price did not move.

What a small team should do this week

Keep the model ID. If dispatch notes, form-to-CRM copies, or intake summaries already call claude-sonnet-5, do not rebuild the workflow to dodge a hike that is no longer scheduled. Teams already routing documents through US Tech Automations workflows will plug this in as a model swap, not a rebuild: same connectors, same human approval step, same log.

Turn on the two discounts you already pay for. Put stable instructions and schema in a prompt-cache breakpoint. Move overnight classification, transcript cleanup, and backfills to the Message Batches API. Set effort to medium or low on jobs a person will edit anyway. Those three changes stack with the freeze; none of them requires a new vendor.

Put a dollar cap next to the rate. The Start tier $500 monthly cap is enough to strand a clinic on the first of the month. Mirror that cap in your own billing alert so a runaway agent cannot spend the freeze’s savings in an afternoon.

Keep a second model on a switch, not in the happy path. Haiku 4.5 at $1/$5 on the pricing table is the in-family cheaper step. luna at $0.20/$1.20 or DeepSeek Flash at $0.22/$0.66 are the out-of-family cheaper steps. The freeze means you are no longer forced to switch; it does not mean you should delete the switch.

Agencies that already compared form-to-CRM automation tools or marketing-agency workflow tools should ask the vendor one question: “Is this invoice on frozen Sonnet 5, on Sonnet 4.6 at $3/$15, or on a seat plan?” If the vendor cannot answer, the freeze never reached you. The same check belongs on executive-assistant automations that quietly call a chat API behind a helpdesk.

For shops that want the model call inside a workflow they can inspect, US Tech Automations documents the platform path as a model step on an existing graph. Agencies mapping form-to-CRM paths on US Tech Automations can keep the same Claude step and change only the model string and the spend alert. That is the whole implementation: ID, cache, batch, cap.

Governance around a cheaper agent

A lower rate makes it easier to run more agents, which is the actual risk. NIST’s AI Risk Management Framework is voluntary guidance for trustworthiness in design, development, use, and evaluation; NIST released the AI RMF on January 26, 2023 and the generative-AI profile NIST AI 600-1 in July 2024 (editorial approval 07-25-2024). The freeze does not replace that work. If a clinic now runs twice as many chart summaries because the unit cost dropped, the failure mode is still a bad summary in a chart, not a misquoted token price.

Price context outside the API table is noisy. According to the U.S. Bureau of Labor Statistics Producer Price Index, final demand prices were unchanged in July 2026 on a seasonally adjusted basis, and the index was +4.7%(p) over the 12 months ended in July 2026. A frozen $2/$10 software line is moving against that backdrop, not with it. That is not a forecast that Anthropic will never reprice again; it is why a lock on a named SKU is operationally useful in a quarter when other producer prices are still up year over year.

Signal vs Speculation

Demonstrated fact (sourced). Claude Sonnet 5 launched June 30, 2026 at $2/$10. Anthropic had published $3/$15 as the standard rate due September 1. On August 10, 2026, Anthropic wrote that the introductory $2/$10 rate is now permanent and that the September 1 $3/$15 schedule no longer applies. Current API docs still show $2/$10 for Sonnet 5, $3/$15 for Sonnet 4.6, $1/$5 for Haiku 4.5, and $5/$25 for Opus 5. Batch is 50% off. Cache hits on Sonnet 5 are $0.20 / MTok. OpenAI luna, DeepSeek Flash, and Gemini 3.8 Flash all list below $2/$10 on the pages cited above. Consumer Pro/Max/Enterprise seat prices are separate from the API freeze. Tokenizer expansion of roughly 1.0–1.35× is a published tradeoff. Retirement is committed not sooner than June 30, 2027.

Our read (12–36 months, not a sourced forecast). If the $2/$10 lock holds, small shops that already standardized on Sonnet 5 will stop treating model cost as a quarterly surprise and start treating token volume as the thing they manage. That usually means more cached prefixes, more batch windows, and more “Haiku or luna for the first pass, Sonnet 5 for the exception” graphs. If Anthropic later launches a Sonnet 5.x at a new list price, the freeze language on this SKU will not automatically cover the successor; read the next changelog the way this one had to be read. Competitive pressure from sub-$1 Flash-class and DeepSeek-class APIs makes a re-hike of Sonnet 5 less likely than a new SKU beside it, but that is a reading of the tables, not a promise. For a 2-truck shop, a 10-person agency, or a solo clinic, the practical 12-month move is: keep Sonnet 5 where quality already cleared, put a spend cap on it, and measure tokens per job so a 1.35× tokenizer does not quietly eat the 50% you just kept.

What we are not claiming. We are not claiming Anthropic will never raise Sonnet 5 again. We are not claiming $2/$10 is the lowest rate a shop can get. We are not claiming Bedrock or Vertex invoices match the first-party table. We are not claiming consumer Pro at $17–$20 absorbed the freeze.

Frequently asked questions

What is the Sonnet 5 Price Freeze?

It is Anthropic’s August 10, 2026 decision to make Claude Sonnet 5’s $2/$10 per million token rate permanent and to cancel the $3/$15 hike that had been scheduled for September 1, as written on the launch changelog. It is a pricing lock on one model ID, not a new model and not a seat-plan discount.

Did Anthropic actually cancel the September hike?

Yes. The August 10, 2026 edit states that introductory pricing of $2 per million input tokens and $10 per million output tokens is now permanent and that the standard $3 / $15 pricing previously set to take effect September 1 no longer applies. The live API price list still shows $2 / $10 for Claude Sonnet 5.

Is frozen Sonnet 5 cheaper than OpenAI or DeepSeek?

No, not on raw list price versus the cheapest SKUs. OpenAI lists gpt-5.6-luna at $0.20 / $1.20 short-context, and DeepSeek lists deepseek-v4-flash off-peak cache-miss input at $0.22. Frozen Sonnet 5 is cheaper than Sonnet 4.6 ($3/$15) and cheaper than the cancelled Sonnet 5 “standard” ($3/$15). Quality, tools, and data-handling still decide whether a swap is worth it.

Does the freeze apply on Amazon Bedrock and Google Cloud?

Anthropic says Claude models are available on Amazon Bedrock and Google Cloud, and that those clouds invoice you. This article does not restate a Bedrock or Vertex dollar figure for Sonnet 5 because those partner pages did not yield a Sonnet 5 cell in the fetch used here. Ask the cloud bill, not this page, for the partner rate.

Will $2/$10 last for the life of the model?

Anthropic’s freeze language is “now permanent” and “no longer applies” for the September 1 $3/$15 schedule, on the launch post. The model overview commits to retirement not sooner than June 30, 2027. Those are not the same promise. A successor SKU can ship at a different list price while Sonnet 5 stays frozen.

How should an HVAC shop, agency, or clinic change its stack?

Do not rebuild. Confirm the production model ID is claude-sonnet-5, cache the stable prompt, batch the work that can wait, set a monthly spend cap, and keep Haiku or a cheaper external model on a fallback route. Clinics that dispatch intake through US Tech Automations can leave the model ID in place and add the cap on the same graph. If a vendor invoices you, ask which SKU and which rate they are passing through.

Do consumer Claude Pro or Max plans change with the freeze?

Seat prices on Claude plans are listed separately (Pro $17 annualized / $20 monthly, Max from $100, Enterprise $20 per seat plus API usage). The freeze text in the launch changelog is about API token rates. A Pro seat that never hits the API is not billed at $2/$10.

Glossary

• Sonnet 5 Price Freeze: Anthropic’s August 10, 2026 lock of Claude Sonnet 5 at $2/$10 per million tokens and cancellation of the September 1 $3/$15 hike.

• MTok: Million tokens, the unit on Anthropic, OpenAI, DeepSeek, and Gemini price tables cited above.

• Prompt caching: API feature that stores a prompt prefix; on Sonnet 5 a hit is $0.20 / MTok versus $2.00 base input.

• Message Batches API: Asynchronous Messages endpoint at 50% of standard rates, limited to 100,000 requests or 256 MB per batch.

• Effort: Request setting (low / medium / high / xhigh / max) that trades token spend for thoroughness; Sonnet 5 defaults to high.

• Context window: Working memory for one request; 1M tokens on Sonnet 5 API, 200k shown on consumer plan cards.

• Tokenizer multiplier: Sonnet 5 footnote that the same text can map to about 1.0–1.35× as many tokens as the prior tokenizer.

• ITPM: Input tokens per minute, a rate-limit bucket; on most Claude models cache reads do not count toward it.

The freeze is a rare case where the vendor took a published increase off the calendar. If Sonnet 5 already clears your quality bar, keep it, meter it, and spend the avoided 50% on logs and a fallback rather than on a rewrite. To see how a locked model ID sits on a graph you can inspect, open the agentic workflow builder or start from the US Tech Automations home page and walk the same dispatch, CRM, or intake path with the frozen rate in the cost column.