This past week an unusual price movement took place in the AI market: while Anthropic makes its cheap rate for Claude Sonnet 5 permanent, DeepSeek is raising its API prices by a factor of up to eleven. For Dutch organisations budgeting for AI tools, this is a good moment to recalibrate the cost structure.
Anthropic freezes Sonnet 5 at introductory price
On 10 August 2026, Anthropic announced that the introductory price of Claude Sonnet 5 will become permanent. Sonnet 5 launched at the end of June for $2 per million input tokens and $10 per million output tokens, with the expectation that the price would rise to $3/$15 from 1 September. That increase will not go ahead.
The official announcement is unambiguous: "We're making Claude Sonnet 5's introductory pricing permanent". The price of $2/$10 per million tokens remains unchanged.
Sonnet 5 is specifically designed for agentic tasks: the model makes plans, uses browser tools and terminals, and works autonomously at a level that previously only larger and more expensive models reached. For marketing and recruitment teams automating workflows, such as content pipelines or CV screening, this rate is competitive with GPT-5.6 Luna ($0.20/$1.20), but with considerably more reasoning capacity.
See our Claude model page for the full comparison overview, including the privacy profile and data processing agreement.
DeepSeek V4: up to 1100% more expensive from 16 August
At the other end of the spectrum: on 16 August 2026, DeepSeek raised the API prices of its V4 models sharply, alongside the introduction of a peak/off-peak tariff structure. Peak hours are defined as 01:00–04:00 and 06:00–10:00 UTC; outside those hours, half rates apply.
The increases vary by model and token type: the lowest is 50%, the highest well over 1100% for specific model combinations during peak hours, according to Qz.com and Emerce. DeepSeek-V4-Flash, which previously cost $0.14/$0.28 per million tokens, is rising too, although DeepSeek is publishing the new rates in stages.
That DeepSeek is raising its prices is notable but not surprising: the model grew rapidly in popularity precisely because of its low costs, and that demand pressure carries infrastructure costs. For teams using DeepSeek at high volumes, such as customer service teams handling large numbers of transactions, recalculation is necessary.
An additional consideration for Dutch organisations: DeepSeek processes data in China and does not fall under the data processing agreement structure that European providers offer. The higher price does not change that profile. On the safety page you can find an overview of which models do offer a data processing agreement.
GPT-5.6 Luna also for free users
On 6 August, OpenAI extended the availability of GPT-5.6 Luna to free users in ChatGPT. Luna, which had previously fallen in price by 80% via the API, is now the default fast model for everyone, even without a paid subscription. This increases the model's visibility, but for organisations working via the API nothing changes in the rates.
What this means for your model choice
The market is now sending two clear signals. On the one hand, efficient models for agentic tasks (Sonnet 5) are being kept cheaper, which makes serious automation more financially attractive. On the other hand, models that ran on extremely thin margins (DeepSeek) may not be able to hold that price at scale.
For organisations budgeting on the basis of current rates: preferably build in a margin of 20–30% as a buffer for announced or unexpected tariff changes. And with every model selection, check whether the privacy profile matches your GDPR obligations; a lower token price does not compensate for a missing data processing agreement.
Sources: Anthropic via X, 10 Aug 2026 · Qz.com, DeepSeek price increase · Emerce, DeepSeek announcement · Dataconomy NL