Within a few days, Google, Deepseek and xAI have released new models that all target the same field of application: code generation and agentic workflows, i.e. AI systems that work through a task independently over many individual steps.What is striking is less the leap in performance than the price movement – in both directions.
Google: new generation after three weeks
Google has presented Gemini 3.7 Flash, only three weeks after its predecessor 3.6 Flash.The model is positioned as the company's most capable working model for code and AI agents.According to Google's own benchmarks, it surpasses Claude Sonnet 5 and GPT-5.6 Terra in code quality, at half the price.
Two limitations are important here.Firstly, these are manufacturer claims, not independent measurements.Secondly, a release interval of three weeks says more about market dynamics than about the maturity of the individual version.Anyone who hard-codes a model choice into an application today may already be working next month with a decision that would better no longer be one.
Deepseek: better performance, significantly higher prices
Deepseek has taken its flagship V4-Pro out of the test phase and, with Harness v0.1, is for the first time releasing its own agent software under the MIT licence – a permissive open-source licence that allows commercial use and modification.For teams building their own agent infrastructure, this is a remarkable building block.
The second piece of news is less pleasing: API prices are rising significantly.Cache hits are particularly affected, becoming six times more expensive.With prompt caching, recurring input parts – for example always the same source files or system instructions – are cached and billed more cheaply than fresh inputs.This is exactly what agentic code workflows live on: an agent reads the same files repeatedly across dozens of steps.
For such applications, the increase in cache costs is therefore the most expensive part of the change, not the nominal token price on the price sheet.Anyone using Deepseek in pipelines should place the actual bill from recent weeks next to the new price structure instead of merely comparing base prices.
xAI: more performance per euro through fewer steps
Grok 4.6 from xAI reaches 61 points in the Artificial Analysis Intelligence Index and is thus on a par with GPT-5.6 Sol and only just behind Claude Opus 5.More interesting than the index value is the information on efficiency: in agentic tasks, Grok 4.6 solves complex workflows in around 53 steps, while Claude Opus 5 requires around 103 steps – at a price more than 60 percent lower.
For project budgets, the number of steps is often more meaningful than the price per million tokens.Every step of an agent generates new inputs and outputs, costs time and increases the probability that the process derails.A model that achieves the result with half the steps lowers costs several times over: in tokens, in runtime and in the effort for error handling.
Something is also moving in the open-source space
At the same time, Ling 3.0 Flash is being talked about as the most intelligent open model in its size class.No reliable details are available at this point, but the reference is relevant: the gap between commercial APIs and openly available models continues to be negotiated in smaller size classes.For scenarios with data protection requirements or self-hosting, it is worth keeping an eye on this branch.
The countermovement: expensive top models are left on the shelf
One figure from another direction is remarkable.Anthropic's flagship Fable 5 is considered the most powerful model on the market, but is hardly purchased by US companies: according to Ramp data, only six percent of current Anthropic tokens are attributable to this model.The obvious interpretation is that a price ceiling for the spread of AI has been reached – at least when the gain in performance and the benefit in everyday work remain abstract.
This fits the rest of the picture: demand is not concentrated on the strongest model in each case, but on the cheapest model that is still sufficient for the specific task.Gemini 3.7 Flash and Grok 4.6 are pushing into exactly this gap with their price arguments.
What this means in practice
A few sober conclusions can be drawn from these reports:
- Keep model selection interchangeable. With release intervals of a few weeks, the provider used should be configurable in applications, not hard-wired.
- Measure costs on your own workload. Manufacturer benchmarks and price lists do not replace a test run with your own repositories, prompts and termination conditions.
- Evaluate caching separately. If cache hits can increase sixfold, the cache strategy is a separate cost item and not a marginal note.
- Count steps, not just tokens. In agentic processes, the number of iterations is a decisive factor in costs and reliability.
- Buy top performance only where it counts. For routine tasks, the cheaper class is usually sufficient; the expensive model remains reserved for difficult special cases.
One realistic caveat: all performance figures mentioned come either from the providers themselves or from aggregated indices.They say little about how a model behaves in a concrete codebase with grown structures, its own conventions and legacy baggage.
This review remains the responsibility of the project team.
For web and software projects, this means above all: model selection is no longer a fundamental decision, but an operational variable that is reviewed at regular intervals – with an abstraction layer in the architecture and a small, repeatable test scenario from your own day-to-day project work. Anyone who monitors costs and quality together can take advantage of price reductions and recognise price increases such as those with cache hits early on. And anyone who uses the appropriate rather than the most powerful model class for each task keeps AI support within budget in the long term.
Sources
- Google's new Gemini 3.7 Flash is said to surpass Claude Sonnet 5 and GPT-5.6 Terra in code quality — The Decoder
- Deepseek releases improved V4-Pro model, raises API prices and makes its agent software open source — The Decoder
- Ling 3.0 Flash is the most intelligent open model of its size — The Decoder
- xAI's Grok 4.6 catches up with the best AI models and significantly undercuts them on price — The Decoder
- AI is becoming too expensive: Fable 5 is a virtual shelf-warmer among enterprise customers — The Decoder