QWE AI Academy Anthropic dropped Claude Opus 4.7 on April 16, 2026, and people are mad. Not at what it can do. At the pricing story everyone’s repeating – incomplete, and production teams are finding out the hard way.
Kept the same per-token price ($5 input / $25 output per million tokens). Every headline: “pricing unchanged.” Buried in the migration docs: the new tokenizer uses up to 35% more tokens for the same input text. Your cost per request just went up. Rate card didn’t.
What Changed (and What Broke)
Opus 4.7 is the first Claude model with high-resolution image support – 2576px vs. the old 1568px cap. More than 3x the pixel count. For reading dashboards, OCR on screenshots, analyzing charts: real improvement. One security research team saw visual accuracy jump from 54.5% to 98.5% when reading dense UI elements.
Coding benchmarks moved. CursorBench went from 58% to 70%. SWE-bench Pro hit 64.3%, beating both GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%). Anthropic’s own tests: 13% resolution lift on a 93-task coding benchmark, solved four tasks that neither Opus 4.6 nor Sonnet 4.6 could handle.
Migrating existing code? Messy.
Three Breaking Changes That Will Stop Your Code
Extended thinking budgets are gone. If your API calls set thinking: {"type": "enabled", "budget_tokens": N}, you get a 400 error. The old way of capping reasoning tokens is dead. Switch to adaptive thinking with the new effort parameter.
# This breaks on Opus 4.7
thinking = {"type": "enabled", "budget_tokens": 32000}
# This works
thinking = {"type": "adaptive"}
output_config = {"effort": "high"}Temperature, top_p, and top_k were removed. Code that tweaks these sampling parameters starts returning 400 errors. Official migration guide: omit them entirely, use prompting to guide behavior. Your workflow relied on dialing creativity up or down via temperature? That control is gone.
Thinking content is hidden by default. If your UI streams Claude’s reasoning to users, they see a long pause plus final answer. The reasoning blocks still exist in the response, but the thinking field is empty unless you set "display": "summarized". Silent change – no error, just broken UX.
The Tokenizer Tax
Same pricing. More tokens.
Opus 4.7 ships with a new tokenizer. Same paragraph of code, same JSON payload, same user message breaks into 1.0x to 1.35x more tokens than it did in 4.6. Upper end – 35% more – shows up most on code, structured data, non-English text.
Request that cost you $0.10 on Opus 4.6? $0.135 on Opus 4.7. One cost analysis: “Anthropic did not raise prices. Your bill may still grow.”
Before migrating production workloads, replay real traffic side by side. Measure the token delta. Don’t trust the 35% ceiling as a flat estimate. Don’t trust 0% either. The only number that matters: what your actual prompts consume.
Caching can absorb some of this. Cache reads are priced at ~10% of standard input rate. Already caching system prompts, tool definitions, or conversation history? Tokenizer change hurts less. Not caching anything? This is a stealth price hike.
The GitHub Copilot Surprise
GitHub rolled out Opus 4.7 to Copilot users with a 7.5x premium request multiplier (promotional pricing through April 30). Opus 4.6 was 3x. More than double the quota cost for what GitHub is positioning as the same price tier. Developers noticed. One GitHub discussion thread: “Rolling out Opus 4.6 and leaving only us with 4.7 at x7.5 (promotional!?!?) is unreasonable.”
GitHub says the multiplier reflects the new tokenizer plus higher reasoning token usage. Community response: unfriendly.
What It’s Good At
Strip away the pricing drama: model does three things better than 4.6.
Long-running agentic tasks. Multiple partners report Opus 4.7 working coherently for hours on problems that used to cause earlier models to give up or loop. Cognition (the team behind Devin): it “pushes through difficult problems that previously caused models to stall.” Notion measured a 66% drop in tool-calling errors on multi-step workflows.
High-resolution vision work. The 3.75MP image cap makes dense screenshots, technical diagrams, data-heavy dashboards legible. Building computer-use agents that need to read small UI text or extract table data from scanned documents? Vision jump is the headline feature.
Stricter instruction following. Cuts both ways. Opus 4.7 follows your prompt more literally than 4.6 did. Ambiguous instructions that 4.6 interpreted loosely get executed exactly as written. Useful for structured tasks where you want precision. Problem if your prompts relied on Claude filling in the gaps.
Effort Levels and the New xhigh Tier
Anthropic added a new xhigh effort level between high and max. Claude Code defaults to xhigh for all plans now. Gives finer control over the reasoning-speed-cost tradeoff.
low / medium: Cost-sensitive, latency-sensitive, or tightly scoped work. Less capable on hard tasks than higher effort levels, but still beats Opus 4.6 at the same effort – sometimes with fewer tokens.
high: Balanced intelligence and cost. Running concurrent sessions or want to spend less without a big quality drop? Start here.
xhigh: Recommended for agentic coding, API design, schema migration, legacy code refactoring. More reasoning, slower, more expensive.
max: Maximum capability. Slowest, most expensive. Use it when the task justifies the cost.
One thing: leave effort at default plus enable adaptive thinking? Opus 4.7 decides per-question how much to think. Simple queries come back fast. Hard ones burn more tokens. You get both. You lose predictability.
Compared to GPT-5.4 and Gemini 3.1 Pro
On directly comparable benchmarks, VentureBeat noted Opus 4.7 leads GPT-5.4 by 7-4. Tight. Where Opus 4.7 wins: agentic coding (SWE-bench), scaled tool use (MCP-Atlas at 77.3% vs GPT’s 73.1%), computer use (OSWorld at 78.0% vs GPT’s 75.0%), financial analysis.
Where it loses: agentic search. BrowseComp dropped from 83.7% on Opus 4.6 to 79.3% on Opus 4.7. GPT-5.4 Pro leads at 89.3%. Workload is web research or autonomous browsing? Test both before switching.
| Benchmark | Opus 4.7 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-bench Pro | 64.3% | 57.7% | 54.2% |
| SWE-bench Verified | 87.6% | – | 80.6% |
| BrowseComp | 79.3% | 89.3% | 85.9% |
| GPQA Diamond | 94.2% | 94.4% | 94.3% |
| OSWorld-Verified | 78.0% | 75.0% | – |
Graduate-level reasoning (GPQA Diamond) is saturated – all three models score within 0.2 percentage points. Real differentiation: applied multi-step work, not raw IQ.
The Mythos Shadow
Anthropic released Opus 4.7 while advertising it’s “less broadly capable” than Claude Mythos Preview, the model they’re keeping locked down. Mythos sits at the top of most benchmark charts – 77.8% on SWE-bench Pro vs. Opus 4.7’s 64.3% – but it’s only available to a handful of enterprise partners for controlled security testing.
Community reaction: sarcastic. Gizmodo’s headline: “Anthropic Releases Claude Opus 4.7 to Remind Everyone How Great Mythos Is.” Bold strategy – promote your new release by telling people it’s not as good as the one they can’t have.
Community Backlash and What It Means
Within 48 hours of launch, the tone on Reddit, Discord, and Hacker News turned negative. Common complaint: Opus 4.7 feels like the pre-nerf version of 4.6 dressed up as a new model, with a tokenizer change functioning as a stealth price increase.
Some of this is vibes. But there’s substance. Users who relied on Extended Thinking toggles to manage cost and latency are now stuck with Adaptive Thinking, which makes the decision for you. Removal was framed as simplification. Power users wanted control, not simplicity.
Removal of temperature/top_p/top_k sampling parameters hit creative workflows hard. Tuning those values to get specific stylistic outputs? That lever is gone. “Use prompting instead” is Anthropic’s answer. Prompting is a blunt instrument compared to direct parameter control.
Backlash overblown? Partly. Rooted in real frustration with API design decisions that choose simplicity over configurability? Absolutely.
What to Do If You’re Already Running Opus 4.6
Update your model ID to claude-opus-4-7. Hard-coded references to 4.6 keep calling the old model.
Audit your prompts. If they say “be brief” or “skip the obvious parts,” Opus 4.7 follows those literally. Re-read any project instructions or system prompts.
Remove temperature, top_p, top_k, and thinking.budget_tokens from API calls. They return 400 errors now.
Stream reasoning to users? Add "display": "summarized" to the thinking config. Otherwise your UI pauses for seconds with no feedback.
Replay a sample of real production traffic through both 4.6 and 4.7. Compare token counts. Don’t migrate at scale until you’ve measured the actual cost delta.
Claude Code users can run /claude-api migrate this project to claude-opus-4-7 to auto-update IDs, parameters, and effort settings. Review the diff before committing.
FAQ
Is Opus 4.7 better than GPT-5.4 for coding?
On SWE-bench and CursorBench, yes. On agentic search and some multilingual benchmarks, no. Your work is coding, agents, or long-running multi-step tasks? Opus 4.7 is stronger. Research-heavy web browsing? GPT-5.4 Pro has the edge.
Why does the same prompt cost more on Opus 4.7 if pricing is unchanged?
New tokenizer breaks text into more tokens – up to 35% more depending on content type (code and structured data see the biggest increase). Per-token price stayed at $5/$25 per million. Number of tokens you’re billed for went up. Effective cost per request rises. Rate card didn’t. This isn’t a bug. It’s how the new tokenizer works. Cost-sensitive? Measure token usage on your actual prompts before migrating production workloads. Caching helps – cache reads are ~10% of input cost. Reused content (system prompts, conversation history, tool definitions) absorbs the tokenizer change better than fresh input. One debugging session that cost $0.10 on 4.6? $0.135 on 4.7. Multiply that by thousands of requests. Your bill grows even though the rate card says “unchanged.”
What happens if I try to use temperature or top_p with Opus 4.7?
400 error. Those parameters were removed. Anthropic’s migration guide says omit them, use prompting to guide model behavior instead. Your workflow depended on tuning temperature for creative vs. deterministic outputs? Rewrite those controls into your system prompt. It’s a real limitation – prompting is less precise than direct sampling control – but there’s no workaround. The parameters are gone. Some users are keeping 4.6 endpoints live specifically for workflows that need temperature control. Not sustainable long-term, but it buys time to rewrite prompts or find alternative approaches.
Pick a small internal task that’s been giving you trouble – complex debugging, dense screenshot analysis, or a multi-step agent workflow. Run it on Opus 4.7 at xhigh effort and see what changes. Vision or reasoning improvements solve something that was brittle before? Migrate deliberately. Feels like the same model with a higher token bill? Stay on 4.6 until the value proposition is clearer.