Tokenmaxxing's bill got big enough to change the price tag.

What mattered this week

Where the next move is

The top of the router is cheap and fast, not flagship.

OpenRouter's live rankings through June 28 put Deepseek V4 Flash first at roughly 630 billion tokens in a single day, with Xiaomi's Mimo V2.5 and OpenRouter's own Owl Alpha close behind. Read it as a surface-specific signal — tokens routed through one marketplace, not global model share — but the pattern is telling: the volume is pooling around fast, low-cost models, not the priciest frontier tiers.

• Treat OpenRouter rank as a list of what to evaluate, not a verdict on what is best; it measures one marketplace's traffic.
• The leaders are cheap-and-fast models, so let that pressure-test whether your agents default to a pricier tier than the task needs.
• Score any swap by accepted output, latency, and retries, not by the headline price-per-million.

The tooling conversation is consolidating around proof, not novelty.

The active building this quarter is less about new model wrappers and more about gateways, traces, evals, retrieval, and token accounting — the parts that make a bill explainable after the fact. That is the same accountability pressure showing up one layer down in the stack.

• Gateways and routers turn model choice into a written policy instead of a per-engineer habit.
• Observability ties spend back to a workflow, an owner, and an accepted artifact.
• Eval and tokenizer tooling keeps a cost cut from quietly becoming a quality regression.

Price your agent runs the way vendors are about to price you.

If the market is moving to per-outcome billing, run your internal numbers the same way. Stop measuring an agent session as one blob; split it into planning, retrieval, edits, tests, and review, and attach the model, tokens, cache hits, and retries to the artifact each step actually produced.

• Define the outcome first — a merged PR, a resolved ticket — then divide total tokens by accepted outcomes for a real unit cost.
• Cap retries and context growth at the start of the run, not after the loop has spent the budget.
• Review your five most expensive runs each week, successes and failures alike, and ask which step you would route cheaper.

Big impressions, thin clicks on the headline term.

A transparency note on our own surface: Search Console shows the query token maxxing pulling about 4,989 impressions but only 57 clicks — roughly a 1.1% click rate at an average position of 7.8. We rank on page one for the term people actually type and are leaving most of the click on the table, which is a title-and-meta problem, not a ranking one.

• Top SEO move this week: rewrite the title and meta for token maxxing to match search intent, since position is fine and click-through is not.
• Router and leaderboard claims in this issue are labeled by scope so a marketplace ranking never reads as global model share.
• The source mix skews to the freshest cost-and-pricing items; older agent-quality context is included as evergreen, not as news.