30 Days With Claude Opus 4.7 in Production: 1M Context Wasn't the Story

I wired Claude Opus 4.7 into QuotaHit on the day it dropped. The 1M token context at standard API pricing is the headline everyone clicked on. The features that actually changed how I ship something else.

This is a 30-day log of what happened inside my own production code — one commit (6b17834 feat(angelina): wire Claude engine via Sonnet 4.6 / Opus 4.7), one voice agent, one set of dispatch-driven workflows. No benchmark theory. What broke, what got easier, and which features the launch blog buried.

The 1M context is the headline. Here's what it actually means in production.

Anthropic confirmed it directly: Opus 4.7 ships a 1M token context window at standard API pricing, with no long-context premium. That last clause is the entire pitch. The other 1M-context Claude releases priced a premium above 200k tokens. This one does not.

What that unlocks in my real code: I can hand the model an entire feature directory — schemas, route handlers, test files, related docs — without compacting. The d table dumps, the OpenAPI spec, the LLD markdown, and the relevant existing route can all sit in one prompt. My PRD → HLD → LLD → CLAUDE.md → Build dispatch loop got noticeably less reduce-summarize-and-pray.

The catch arrives on your invoice. The new tokenizer uses roughly 1x to 1.35x as many tokens as Opus 4.6 for the same English text. Anthropic flagged this in the breaking changes section, and it shows up immediately if you run a workload through both models. "Free 1M context" is true on the per-token rate. The same task still costs more if it tokenizes 30% heavier. Reset your cost dashboards before you ship anything with this assumption baked in.

2576px vision is the silent killer feature.

Maximum image resolution went from 1568px / 1.15MP to 2576px / 3.75MP. That's not a marketing bump — that's the floor at which screenshot agents stop guessing.

If you're building anything that reads UIs, the difference is concrete. Computer-use agents, document review (.docx / .pptx per Anthropic's own examples), chart and figure analysis with pixel-level transcription — all of these previously had to downsample or tile to fit. Now they do not. Coordinate mapping is also 1:1 with actual pixels, so no more scale-factor math when the model points at a UI element.

For QuotaHit's roadmap, this changes how I'll build the "show me the lead's IndiaMART screenshot and tell me what's missing" feature when it ships. The vision call no longer eats accuracy at the pre-processing step.

Task budgets are a better mental model for agent cost control.

This one's beta and easy to miss. With the header task-budgets-2026-03-13, you can pass an advisory task_budget to the model — a running countdown across the full agentic loop, including thinking, tool calls, tool results, and final output.

• -Critical distinction the docs spell out:
• -max_tokens is a hard per-request cap. The model never sees it.
• -task_budget is advisory. The model sees the countdown and self-moderates.

For QuotaHit's voice agent, that's the difference between "agent gets cut off mid-tool-call" and "agent wraps up the call gracefully when it senses the budget tightening." Minimum value is 20k tokens. Below that, the model may refuse the task or finish it poorly.

I haven't promoted this from beta to default yet. But for any agent loop where cost matters more than open-ended quality, this is the lever that's been missing from the platform.

The breaking changes the launch blog buried.

Three breaking API changes that aren't on anyone's Twitter feed:

1. temperature, top_p, top_k are removed. Setting any of them on Opus 4.7 returns a 400. If you had temperature: 0 for "determinism" — note that it never actually guaranteed identical outputs — you now have to use prompting alone. I had to grep my entire Angelina codebase for temperature and strip it.

2. Extended thinking budgets are gone. The old thinking: {type: "enabled", budget_tokens: N} pattern returns a 400. Adaptive thinking is the only mode now, and it's off by default. If you were depending on thinking being implicitly on, your prompts will degrade silently until you set thinking: {type: "adaptive"}.

3. Thinking content is omitted by default. Thinking blocks still appear in the response stream, but their thinking field is empty unless the caller explicitly opts in with display: "summarized". If your UI streams reasoning to users, this will look like a long pause before output begins. Mine did.

None of these crash production. All of them silently degrade behavior. Re-baseline anything you migrate.

The tone shift nobody's talking about.

Anthropic put this one in plain English: "More direct, opinionated tone with less validation-forward phrasing and fewer emoji than Claude Opus 4.6's warmer style." And: "More literal instruction following, particularly at lower effort levels. The model will not silently generalize an instruction from one item to another."

For product builders, that line matters. Prompts that worked on 4.6 because the model was inferring intent — "be helpful and friendly" — get colder on 4.7. The model also no longer extrapolates from one example to the next without being told to. Your few-shot scaffolding needs to be explicit, not suggestive.

I had to give Angelina's voice prompt an extra paragraph reaffirming the persona ("warm, Indian-English-fluent, asks back-and-forth questions"). On 4.6 that personality emerged. On 4.7 you have to ask for it directly.

My updated routing across all 4 products.

The economics have shifted enough that the routing math is different now. Where I sit on 2026-05-27:

• -Haiku 4.5 — high-volume, low-stakes (intent classification, simple summarization, transcript cleanup in IndianWhisper).
• -Sonnet 4.6 — general code work, conversation layers, anything with a tight token budget where Opus is overkill.
• -Opus 4.7 — complex agentic tasks, anything that needs long-context reasoning (codebase audits, multi-file refactors), high-effort coding.

Inside QuotaHit specifically, the M11-E01 commit wired both Sonnet 4.6 and Opus 4.7 behind Angelina's engine layer with a routing call. Conversation turn → Sonnet 4.6. Multi-step reasoning or tool-orchestrated action → Opus 4.7. The cost delta justifies itself when the agent doesn't have to retry.

What to do this week if you haven't migrated.

If you've been running Sonnet 4.6 as your default and treating Opus as the special-occasion model, three things to actually test:

• -Run your single hardest agentic loop on Opus 4.7 with xhigh effort. That's the new effort level designed for coding and agent work. Compare to what you have. If the first-pass success rate jumps, the cost delta probably pays for itself.
• -Audit your codebase for temperature, top_p, top_k, and old thinking budgets. Strip them before you swap in the new model ID.
• -Bump your token budget dashboards by ~30% until you know how the new tokenizer hits your specific workload. Better to over-budget and under-spend than the reverse.

The 1M context will be the line in the headline. The high-res vision, task budgets, and stripped sampling parameters are the line in your migration PR.

I'll keep logging what shifts as I push more dispatches against 4.7 in the next 30 days.

— Dhruv, AI Builder