Claude Opus 4.8 Raises the Stakes for AI Token Management

Anthropic released Claude Opus 4.8 on May 28, 2026, pitching it as the company’s most capable model yet. Within days, the release had become something else too: A case study in how easy it now is to spend far more than intended on a single AI task. The new features genuinely expand what Claude can do on its own, but they also expand how fast a token bill can grow.

New capabilities, same sticker price

Headline pricing didn’t move. Opus 4.8 still runs at $5 per million input tokens and $25 per million output tokens, the same rate held since Opus 4.5. What changed sits underneath that number. Fast Mode, running at roughly 2.5x standard speed, is now priced three times lower than on previous Opus versions, and a new effort control dial lets developers set how much the model reasons, with tiers from high up to max.

The bigger addition is dynamic workflows. Inside Claude Code, the model can now plan a task and fan it out across hundreds of parallel subagents in a single session, tackling something like a code migration across hundreds of thousands of lines from kickoff to merge.

Subagents multiply the token meter, not the price ceiling

Each subagent runs its own token meter with no scale discount, so the cost rises in direct proportion to how ambitious the task is. A widely circulated developer benchmark showed the gap: Opus 4.8 at max effort reportedly burned 16.5 million tokens and $17.26 on a mid-size Cursor ticket that GPT-5.5 finished using 5.9 million tokens for $5.57. Same task, roughly triple the cost.

This pattern predates the release. Research from Ida Silfverskiöld found that an unoptimized agent handling 100 messages a day could cost around $2,490 a month, roughly 25 times more than the same agent tuned properly. Opus 4.8 just raises the ceiling on that same dynamic.

The half-billion-dollar claim nobody can confirm

Part of the conversation has been driven by an unverified claim. Axios reported an AI consultant’s account of a client that allegedly spent roughly half a billion dollars on Claude in a single month after failing to set usage limits on employee licenses. The figure racked up more than 29 million views on a single post, even as several commentators, including the journalist who first reported it, doubted the number. True or not, it landed at a moment when companies were already nervous about AI costs heading into earnings season.

Companies are quietly killing their token leaderboards

Some companies had started treating heavy token use as a badge of being AI-forward, a habit nicknamed tokenmaxxing. That’s now looking like a liability. Amazon reportedly pulled an internal leaderboard ranking employees by token usage after staff started gaming it, and Meta shut down a similar leaderboard. Axios also reported that Microsoft scaled back internal Claude Code licenses partly over cost, with Uber’s COO calling the spend harder to justify than before. These are a handful of named examples, not a market-wide survey, but they point in the same direction: raw token volume stopped correlating with anything companies actually wanted.

The fix: Routing, not retreat

More disciplined teams are routing rather than retreating. Factory, an AI coding platform, routes each query to the cheapest model that can handle it, with its CEO noting many tasks don’t need Opus-level reasoning; the company reports open-model usage tripling relative to closed models in the past month. Cursor’s Composer 2.5 has reportedly matched Opus 4.7 and GPT-5.5 on benchmarks at a fraction of the cost, giving routers a credible, cheaper fallback. Infrastructure is catching up, too: the llm-d project, backed by IBM Research, Red Hat, Google Cloud, NVIDIA, CoreWeave, Hugging Face, and Mistral AI, is built to run any model on any accelerator in any cloud, cutting the lock-in that made switching providers risky.

The takeaway

Opus 4.8 is a stronger model than its predecessor, and also a clear signal that capability and cost are now rising together. The lesson isn’t that AI spending is inherently out of control. It’s that using it carelessly now costs enough to notice, and the more sustainable path is matching tasks to the cheapest model that can do them, using effort controls deliberately, and holding the people closest to the work accountable for what it costs.