How Companies Are Becoming AI Token Efficient

May 4, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
June 04, 2026

HEADLINES

ChatGPT Hits a Billion Users

ChatGPT has officially crossed a billion monthly active users, according to new estimates from Sensor Tower based on May data — making it the fastest app to reach that milestone in history at three and a half years, well ahead of TikTok's five years and YouTube and Instagram's eight. The five-month delay from OpenAI's end-of-2025 target turned out not to matter much: the April WSJ story framing it as a growth plateau was already out of date by the time it published, with OpenAI well into their resurgence by then. Around 12% of the global population now logs into ChatGPT monthly, a scale nothing else in the industry comes close to. Meanwhile, Claude's 640% user growth over the past year is striking — but still only puts them at 56 million monthly active users, about 5% of ChatGPT's base, a gap that makes Anthropic's revenue lead all the more impressive.

Bots Overtake Humans in Web Traffic

For the first time in internet history, bots and agents now account for the majority of web traffic. Cloudflare's data puts bot traffic at 57.5% of what flows through their service, driven by AI data scrapers and the rapid growth of web agents. The downstream effects are real: ad revenue is falling and malicious automated traffic is surging — Cloudflare now classifies 37% of automated traffic as "bad bots" that ignore robots.txt. What's striking is that Cloudflare CEO Matthew Prince predicted this would happen by 2027; he took to X on Wednesday to note it arrived considerably ahead of schedule.

Meta's Business Agent: OpenClaw for Small Business

Meta launched a new business-focused agent at the WhatsApp "Conversations" conference in London, and the framing matters: this isn't really an enterprise play in the traditional sense. It's targeted at the bakery in São Paulo and the clothing shop in Birmingham — the 200 million businesses already running on WhatsApp. The agent automates appointment booking, sales, and eventually market research and calendar management, and will expand into a broader platform with connectors for Shopify, Zendesk, and hundreds of other non-Meta tools. The agent starts free before shifting to a paid subscription. Separately, The Information reports Meta is also planning a consumer "Hatch" agent that could run up to $200 a month and include an integrated vibecoding platform.

MAIN STORY

How Companies Are Becoming AI Token Efficient

The AI story of the moment isn't a new model or a new benchmark — it's the dawning realization across the industry that token consumption has become a genuine cost constraint. Sam Altman said at OpenAI's enterprise event this week that AI costs, which "never came up" at the start of the year, have become a "huge issue" for companies. The shift from assisted AI to deploying fleets of agents has dramatically increased token consumption, while the infrastructure to supply those tokens is still being built out — meaning we're likely living with some form of token scarcity for the better part of the next five years. The question now isn't whether to think about token efficiency, but how.

TIME TO GET EFFICIENT

Efficiency Is Now a Benchmark Metric
Token counts are joining raw scores on leaderboards — and changing the picture.
The clearest sign of the shift is in benchmarking. For most of AI's history, state-of-the-art meant the highest raw score, full stop. That's changing. Artificial Analysis's most important chart is now their intelligence versus output tokens quadrant — and on that chart, Claude Opus 4.8 paints a notably different picture than on the leaderboard alone. While Opus 4.8 edges above GPT-5.5 on the intelligence index, it burns roughly 80-90% more tokens to get there, placing it well outside the most attractive quadrant of high performance and high efficiency. Gemini 3.5 Flash tells a similar story: higher intelligence than its predecessor, but at more than five times the cost to run. Perplexity CEO Aravind Srinivas put it cleanly on CNBC: the company that can deliver the most "token value per watt per user" — balancing accuracy, latency, cost, privacy, and intelligence — is the one that wins long term.

Microsoft Puts Token Usage on the Model Card
Average token usage is now a first-class metric, and Microsoft wants credit for it.
Microsoft is betting that frontier-tuned, token-efficient models are a competitive wedge. VC Tomasz Tunguz flagged that Microsoft has added an "average token usage" column to every benchmark table in their system cards — a move he predicts will become standard across the industry. The numbers are striking: MAI-Code-1-Flash hits 71.6 on SWE-Bench Verified while using a third of the tokens Claude Haiku 4.5 burns. And in a customer example, when tuned for McKinsey's specific tasks, the Microsoft model outperformed GPT-5.5 on quality while costing ten times less. The argument is that task-specific tuning produces models that are simultaneously smarter and cheaper for the work that actually matters — a different axis of competition than raw benchmark scores.

Cursor Composer 2.5
Training your own model on one task turns out to be a powerful efficiency play.
Cursor's Composer 2.5 is the clearest example yet of the agent-layer efficiency strategy: take Kimi K2.5, train it hard on coding and nothing else, and end up with a model that matches frontier performance on coding tasks at radically lower token cost. By not trying to hill-climb every benchmark, Cursor got a much more efficient model for the specific work their users actually do. As Gergely Orosz observed, the company with the cheapest coding model at good-enough quality is in a very strong position as developers and enterprises become more price-sensitive.

Harvey's Multi-Model Routing Results
Smart routing beat brute force — in production, on one of enterprise AI's hardest benchmarks.
Legal AI firm Harvey published results from a hybrid routing experiment: using GLM 5.1 as a primary worker that routes to Opus 4.7 as an advisor only when needed — an average of 0.83 times per task. The hybrid setup outperformed Opus alone on both quality and cost. Harvey also found that post-training Kimi's K2.6 model on legal tasks moved it ahead of Opus on their legal agent benchmark at 11 times lower cost. Patrick Oshag's summary cuts to the point: using the most expensive model for every task isn't a quality strategy, it's a laziness tax.

Factory Router
"You wouldn't have Messi play goalie."
Software development company Factory launched Factory Router this week — a product that automatically picks the right model for every task, with the stated goal of maintaining frontier performance while cutting costs by 20-25%. Their framing captures the new consensus well: a higher token bill doesn't mean more work is getting done, and routing routine tasks to the most expensive model just burns budget.

Perplexity's Hybrid Agentic Inference
Split tasks between local hardware and the cloud — automatically, based on what each sub-task actually needs.
Perplexity demonstrated a hybrid inference system at Computex that breaks agentic tasks into components, routes them to sub-agents, and automatically determines which need cloud inference versus local hardware. The system ran on an Intel Core Ultra 3 consumer device. The practical upshot is an orchestrator that balances intelligence, cost, latency, and privacy in a single automated workflow — including the ability to identify sensitive data and keep it off the cloud entirely.

DeepSeek Tops Ramp's Trending Vendor List
Companies want cheap AI badly enough to route US data through Chinese servers.
Ramp's June trending software vendor report put DeepSeek at the top — a development that surprised even Ramp's own lead economist Ara Kharazian, who had predicted companies would shift toward cheaper models but didn't expect American firms to actually turn to the Chinese competitor directly. Three open-source model service providers also made the list. The takeaway isn't just cost pressure — it's that DeepSeek has become synonymous with cheap AI in a way that Google's Gemma 4, for all its efficiency gains, simply hasn't. That's a marketing problem as much as a technical one.

Arvind Jain's Token Architecture Framework
Token spend is an architecture problem, not just a model selection problem.
Glean CEO Arvind Jain published an essay this week that offers the clearest framework yet for thinking about enterprise token efficiency. He identifies four architectural levers: context quality (too much conflicting or irrelevant context burns tokens before the task even starts); model routing (the goal isn't small models everywhere, but the right intelligence level for each job); continual learning (systems that remember prior successful executions don't have to pay the exploratory cost again and again); and harness design. The central argument is that the companies treating token efficiency as a model-selection problem are solving for the wrong thing — it's an architecture problem all the way down.