Claude Opus 4.7 raises token burn 1.0-1.35x as users report refusals

• Anthropic's launch summary says Claude Opus 4.7 ships at the same list price as 4.6, but with a new tokenizer that maps inputs to 1.0 to 1.35x more tokens and a new xhigh effort tier.
• According to Boris Cherny's launch thread, Claude Code now defaults to xhigh effort for 4.7, while Boris Cherny's rate-limit note says Anthropic raised subscriber limits to offset the extra thinking tokens.
• Cost complaints showed up fast: ozansihay's usage-limit screenshot shows a shared Claude and Claude Code cap already exhausted, while AIandDesign's post says one Claude Design session burned 67% of a $100 Max plan.
• Steering complaints also landed immediately. kaigani's refusal report says Opus 4.7 rejected a story-outline scoring task on safety grounds, and om_patel5's Reddit complaint screenshot collects claims about ignored preferences, skipped search, and fabricated search behavior.
• A separate thread points to a practical workaround inside Claude Code: heyrimsha's benchmark post and hasantoxr's benchmark numbers say the WOZCODE plugin cut one Opus 4.7 task from $14.38 to $6.61 by reducing tool round-trips.

Anthropic's own announcement pitches 4.7 as better for interfaces, slides, and docs, while Boris Cherny's Claude Code tips thread quietly spells out the new defaults. The useful weirdness lives in the gap between those two surfaces: a permissions doc, a very large Hacker News thread, and creator posts showing that a single design or coding session can chew through limits faster than the spec sheet suggests.

Anthropic positioned 4.7 as a stronger Opus for hard software work and polished deliverables, specifically naming interfaces, slides, and docs in the official launch post. The headline price stayed flat at $5 per million input tokens and $25 per million output tokens, but the usage math did not.

According to Boris Cherny's launch thread, Claude Code got three practical changes alongside the model swap:

• xhigh, a new effort level between high and max
• auto mode, which routes permission checks through a classifier so long runs need less babysitting
• more explicit advice to manage spend with effort, task budgets, or brevity prompting

Adaptive thinking also changed the control surface in the consumer app. QualitativAi's screenshot shows Opus 4.7 paired with an "Adaptive thinking" toggle, while Cherny's best-practices thread says Claude Code users should think in effort levels, not old thinking budgets.

The buried caveat in this launch is simple: flat list pricing does not mean flat session cost. Anthropic's launch summary says the new tokenizer maps the same input to 1.0 to 1.35x more tokens, and Boris Cherny's rollout note adds that Claude Code now defaults to xhigh, which pushes reasoning higher unless users dial it back.

That combo showed up as immediate limit pain in both coding and design posts. AIandDesign's complaint says a single Claude Design task consumed 67% of a $100 Max plan, while fresh HN discussion says commenters were already telling each other to pin effort and task budgets if they wanted predictable spend.

Anthropic did respond on limits. Boris Cherny, Claude Code lead at Anthropic, wrote in his rate-limit post that Opus 4.7 uses more thinking tokens and that subscriber limits were increased to compensate, and later amplified a fix for long-context limit accounting in his repost of the ClaudeDevs fix. But the same HN thread still surfaced downstream pricing fallout, including a comment summarized in the core HN recap claiming GitHub Copilot was charging 2.5x more for Opus 4.7 prompts.

The other fast-moving complaint is steerability. kaigani's post says Opus 4.7 rejected a request to score a story outline on safety grounds, and the fresh HN recap in meetpateltech's update highlights a separate report where a previously normal debugging session suddenly tripped usage-policy blocks and burned the day's allotment.

The claims around preference-following are sharper, but also more anecdotal. om_patel5's screenshot of a Reddit complaint lists four alleged regressions: ignored custom preferences, skipped or fabricated web search behavior, unsolicited editorial refusals, and weaker results in warm sessions than cold ones. HN commenters were less certain about a broad model downgrade. According to one HN reply, at least some of the weirdness may be confounded by Anthropic changing Claude Code defaults at the same time.

Anthropic's creative pitch for 4.7 has some support in the early hands-on posts, but only for certain formats. gregisenberg's unscripted test scored Claude Design highly for wireframing, mobile app design, and deck research, while giving video creation a much weaker 4.5 out of 10.

Those first reports line up with the launch framing in the core HN summary, which says Anthropic is explicitly aiming at interfaces, slides, and docs. They also line up with the complaints: _VVSVS's post says even simple layout nudges can burn through credits, and AIandDesign's post suggests the cost ceiling becomes visible faster in design work than in one-shot chat tasks.

The most concrete counter-signal in the evidence set is not a model defense, it is a tooling hack. heyrimsha's benchmark post and hasantoxr's benchmark numbers both compare the same Opus 4.7 task in stock Claude Code versus the WOZCODE plugin.

On that one task, the deltas were:

• Cost: $14.38 to $6.61
• Turns: 236 to 76
• Duration: 22m 44s to 15m 46s
• Output tokens: 102k to 77.4k

The explanation in heyrimsha's follow-up and hasantoxr's tool-round-trip post is that Claude Code's default file and search tools can return small chunks, forcing the model to keep asking for more context. WOZCODE claims to cut that re-asking loop by returning more useful context per call. Installation is also unusually lightweight for this kind of claim: hasantoxr's install post says it is just a Claude Code plugin, added with /plugin marketplace add WithWoz/wozcode-plugin and /plugin install woz@wozcode-marketplace, with code staying on the local machine and Anthropic's API path.