The Models Are Getting Cheaper. Enterprise AI Bills Keep Rising.
On September 14, 2026, the temporary 50% boost Anthropic added to its weekly rate limits in May runs out. In its place comes a smaller, permanent increase of 25% over the original baseline.
Put the baseline at 100. Before May, users had 100. For the last four months they’ve had 150. From now on they get 125. That’s 25% more than the original baseline, and 17% less than what teams have actually been using. Anthropic admitted the cut in its announcement, which went up, came down, and went back up with a clarification.
Anthropic had extended the boost through mid-July, then to July 19, August 19, and August 31, each time hinting it might stay. Teams built that capacity into their weekly work. After four months and five extensions, the extra allowance was simply how people worked, whatever the label said.
What this shows is how exposed you are when you depend on one vendor. If a company’s daily knowledge work runs through a single closed AI provider, that provider sets its real unit costs, its throughput and how far its internal AI workflows can grow. A plan update turns into a cost and capacity problem for finance almost overnight.
Premium models have become the costly default for everyday work
Open models are getting better quickly, and that’s driving down the price of the underlying intelligence. On a lot of everyday knowledge work they come close enough to frontier models that the price gap deserves a hard look at procurement time.
Enterprise AI bills still go up, because companies are buying more than intelligence. They’re also buying a set of defaults: the model employees use for every request, the vendor their workflows rely on, the interface where their work piles up, and pricing they don’t control.
The most expensive default is sending every task to the best model available. A meeting summary and a high-stakes legal analysis deserve different budgets. So does a routine research brief, which often runs on the priciest model only because that model is the default in the interface. A lot of enterprise AI still works like this, with one expensive model applied to everything. As usage grows, the bill grows faster, and everyday workflows lean harder on a single vendor.
Claude Sonnet 5 is a strong model, and some work really does call for premium capacity. Companies still have to decide whether it should take every piece of knowledge work by default.
DeepSeek V4 Flash costs up to 23x less while performing within a few points of Claude Sonnet 5
The comparison that matters most is DeepSeek V4 Flash against Claude Sonnet 5, the model many teams use day to day in Claude.ai and the Claude API.
According to Artificial Analysis, DeepSeek V4 Flash costs $0.44 per million input tokens and $1.32 per million output tokens. Claude Sonnet 5 costs $2.00 and $10.00 for the same. That makes DeepSeek about 4.5x cheaper on input and 7.6x cheaper on output. Per task the gap reaches about 23x: $0.22 against $5.09.
On quality they’re much closer. Sonnet 5 leads the Artificial Analysis Intelligence Index 38 to 35, a real but small lead. It scores 1501 on GDPval-AA v2 to DeepSeek’s 1456, and 1355 on AA-Briefcase to DeepSeek’s 1257. Sonnet is the smarter model on paper, but DeepSeek lands within a few points on the work companies actually do, at a fraction of the price.
A sensible AI strategy picks the model to fit the task. For most organizations, the share of work that needs a premium model is “less than we currently assume.” Cheaper models can write a solid first draft, pull together the right internal material and prepare structured briefings when they have the right context. Running all of that through a premium closed model spends money without giving you better control over quality.
Company context helps lower-cost models handle more enterprise work
Context matters more than anything else in enterprise AI, and model comparisons usually ignore it. A model doesn’t know how your company approves a renewal, how your marketing team describes the product, which documents are current or where your regulatory team draws the line. Without that, even a very capable model has to guess. Give a cheaper model the right context and it gets much better at the work your company cares about.
Most enterprise AI tasks come down to applying what the company already knows, reliably: approved messaging, team playbooks, past decisions and corrections, the current systems of record, and the feedback that shows an agent what good work looks like. When every agent can read that context, and a correction in one conversation sticks for the next one, people can write shorter instructions. Agents can also take on harder multi-step work and produce more consistent results across a team.
Context also makes it less painful to change models. When the value lives in individual prompts and one vendor’s interface, moving to a different model means teams have to rebuild how they work. If context, workflows, governance and agents sit in a platform layer above the model, a company can pick the best model for each task without starting over. Its advantage then comes from its own knowledge, not from whatever a lab’s pricing page says this month.
A model-independent architecture reduces costs and vendor lock-in
Relying on a closed vendor puts the power on their side. Employees adopt the tool, workflows come to depend on it, and usage climbs. Then the vendor can change pricing, packaging, access or capacity, and the customer’s only alternative is the cost of rebuilding the work somewhere else. That’s lock-in, even when the change is announced as a plan improvement.
You can keep using Claude and other frontier models and still choose a different default for everyday work. An AI architecture that lasts sends each task to the right model: cheaper open models for high-volume routine work, premium models for the cases where better quality is worth the extra cost.
In this setup, company context, documents, workflows, feedback and permissions stay independent of the model underneath. Employees get one interface they’ll actually use, which cuts down on scattered tools, shadow AI and spending nobody can manage. Leaders can see usage, cost and how agents are performing, so adoption keeps building across the company instead of starting over with every new hire.
elvex is built to be that platform layer. DeepSeek V4 Flash is available on elvex today, and V4.1-Flash is coming shortly. Beyond model access, elvex gives you company context that builds up over time, one interface employees actually use, and agents and workflows that don’t depend on any provider’s pricing or roadmap. Rollout is organized around how the company really works, not around a stack of licenses.
Review AI capacity before the next vendor change forces the issue
Look at how you buy AI capacity now, while you still have choices, and before the next rate-limit update, packaging change or price increase turns it into an emergency.
On the tasks that make up everyday enterprise work, open models are within a few points of the frontier workhorse, at a fraction of the cost. With the right context, workflows and governance around them, they’re worth far more than their benchmark scores suggest.
Anthropic’s rate-limit change gives companies a concrete reason to revisit their defaults. Paying premium prices for every task, and letting one vendor decide the cost and capacity of company-wide AI, is no longer the only option.