Claude Opus 4.7 with 1M Context for Coding: What Developers Should Test First

Image: Anthropic News, used for editorial coverage of the Opus 4.7 release.

What happened

Anthropic released Claude Opus 4.7 on 16 April 2026 with the 1M-token context window inherited from the broader Claude 4.6 family for Claude Code workflows ¹ . The 1M tier itself went generally available on Claude Opus 4.6 and Sonnet 4.6 on 13 March 2026 (Sonnet 4.5 had carried 1M as a beta from September 2025 until that beta retired on 30 April 2026); Opus 4.7 inherits the 1M ceiling rather than introducing it, so the headline capability shift on Code is the model upgrade plus tokenizer change rather than a context-window debut ² . The longer window sits behind the Code surface on Pro, Max, Team, and Enterprise plans across the 4.6 / 4.7 family; on Pro, the 1M window is opt-in via the /extra-usage command in Claude Code, and Max, Team, and Enterprise plans get 1M without the opt-in step ⁸ .

For developers running long-context refactors, multi-repo migrations, full-codebase audits, and framework upgrades, this is the most material capability shift on the Code surface since Sonnet 4.5 shipped, per the aggregated launch-coverage consensus. Pro is $20 USD per month, which lands at approximately ₹2,240 delivered to a card after 18% GST and forex markup ³ , billed via local cards or international rails through claude.com. Verify the geo-render at checkout; prices fluctuate, so confirm before purchase as of 5 May 2026.

The recommendation is straightforward. Test the 1M window on a real refactor, not a synthetic context-recall benchmark. Pick a multi-file change in a codebase already in production, hand the whole repo to Claude Code, and measure two things: how often the model finds load-bearing context outside the file you would otherwise have hand-selected, and how the wall-clock time compares against the equivalent agentic flow at 200k. Synthetic benchmarks understate both ends.

What is actually new

The 200k-token cap was the default ceiling on Claude’s coding workflows for most of the 4.x family until 13 March 2026, when Anthropic shipped 1M context as generally available on Claude Opus 4.6 and Sonnet 4.6 (Sonnet 4.5 had carried 1M as a beta from September 2025 until that beta retired on 30 April 2026). Opus 4.7 inherits the 1M ceiling from the 4.6 family it succeeds. At roughly 150,000 lines of source code in typical Python or TypeScript density, 200k tokens forced the developer into one of two patterns: hand-curate a sub-tree of files for the model to load, or let an agent loop over grep and read calls until it finds the relevant context on its own. Both patterns leak time. The 1M ceiling on Opus 4.6 / 4.7 and Sonnet 4.6 removes that constraint for codebases under that size.

Opus 4.7 also ships with notable benchmark gains on coding evals (per Anthropic’s published benchmark chart on the launch announcement): on CursorBench Anthropic reports 70% for Opus 4.7 versus 58% for Opus 4.6, with Rakuten-SWE-Bench noted as resolving “3x more production tasks than Opus 4.6”, and SWE-bench Verified / Pro / Multilingual flagged with memorization-screen caveats in the release notes ⁹ . Opus 4.7 also ships an updated tokenizer where 1M tokens equals roughly 555k words / 2.5M unicode characters versus Sonnet 4.6’s roughly 750k words / 3.4M characters at the same 1M ceiling — a token-density penalty of roughly 1.3-1.35x more tokens per file on code-heavy contexts (Python, JSON, TypeScript), so the LOC-to-tokens math under the 1M ceiling does not extrapolate cleanly between Opus 4.7 and Sonnet 4.6. For a developer choosing whether to pay the Opus premium over Sonnet, the CursorBench gain is decision-relevant alongside the tokenizer change.

The 1M window changes the cost-curve at one end. A typical mid-sized SaaS or fintech monorepo, call it 400,000 to 700,000 lines across services, infrastructure, and shared packages, now fits in a single context with room left over for the conversation. That does not mean the model will reliably recall every line under that budget; long-context recall degrades in ways that are workload-specific and not yet well-characterised at the 1M scale. It does mean the developer no longer has to pre-select context for the model to even attempt the task.

Anthropic’s product framing for Opus 4.7 positions the longer window as Code-surface-first because that is where the cap was binding hardest. The consumer chat surface rarely needs more than 200k for a single conversation, and serving a 1M context inference at chat-grade latency is a different operational problem from serving it inside a long-running Code session where the wall-clock budget is measured in minutes, not seconds.

Image: Claude Code product page, used for editorial coverage of the agentic coding surface where the longer window is exposed.

What this unlocks for Indian developers

Three concrete patterns become tractable that previously required heavy pre-selection.

The first is multi-repo refactors. Migrating a Java monolith to a service-mesh layout, or moving a Next.js 13 codebase to Next.js 15, has historically meant hand-curating the relevant files and feeding them to the model in batches. The 1M window holds the whole monolith plus the target framework’s documentation in one context. The agent can cross-reference call sites without losing state between turns.

The second is full-codebase audits. Security audits, dependency-graph reviews, and pre-IPO code-quality passes that Indian compliance teams run before product launches are bottlenecked by the same context cap. A 1M window holds enough surface area to pattern-match across an entire repository without losing earlier findings as new files load.

The third is framework upgrades. The “upgrade Spring Boot 2.x to 3.x across all 47 microservices” task that occupies an Indian platform team for two weeks of senior-engineer time is the canonical fit. Each service is small; the cross-service consistency is the load-bearing concern; the 200k cap forced sequential per-service passes. The 1M window allows a single planning pass across all 47 services before any code changes ship.

None of these are zero-effort tasks at 1M either. The model still hallucinates, still misses context that lives outside the repo (infrastructure config in a separate Terraform repo, runbook PDFs, internal Slack threads), and still produces edits that need human review. The 1M window removes a structural ceiling, not the engineering work below it.

What it does not change

The consumer Claude.ai chat interface keeps its 200k context cap for now, per the model documentation accessed at draft time ⁴ . Free-tier Claude is unaffected; the 1M window is a Pro and Max-tier capability inside the Code product specifically.

Output token caps remain the binding constraint on long-form generation. A 1M input context does not imply a 1M output budget; the model’s per-turn output ceiling is unchanged in the documented model card ⁵ and remains in the low-tens-of-thousands of tokens per response. Long refactors still need to be chunked into multiple turns; the 1M window changes the planning-pass ceiling, not the patch-emit ceiling.

Latency increases meaningfully at the longer window. A 1M-token inference is slower than a 200k inference at the same model tier, and the wall-clock difference matters for interactive developer flows. Indian developers on residential broadband should expect the per-turn round-trip to climb from seconds to tens of seconds at the upper end of the new window. That is a tradeoff developers should measure on their own connection before committing a sprint plan to it.

The 1M context is also not a substitute for retrieval. Codebases that grow past 1M tokens (rare in single-product Indian companies, common in platform-engineering teams at larger SaaS firms) still need RAG or sub-tree selection. The new ceiling is higher; it is not infinite.

Pricing for India

Claude.com displays USD pricing; the geo-render varies and Indian users may see USD or local-currency conversions at checkout. As of 5 May 2026, Pro is $20 USD per month, which lands at approximately ₹2,240 delivered to an Indian card after 18% GST and forex markup ⁶ . Max starts at $100 USD per month (5x usage tier), which lands at roughly ₹11,200 delivered after the same GST and forex layers; Max20, the 20x usage tier, is $200 USD per month or roughly ₹22,400 delivered. Prices fluctuate, so verify before purchase. Both tiers accept international credit and debit cards at checkout; UPI support is documented inconsistently across third-party reporting and could not be confirmed against a primary source at draft time, so verify in-app.

On the question of which tier unlocks 1M, the picture is more nuanced than the sticker prices suggest. Pro can access the 1M window, but only after the developer types /extra-usage in the Claude Code CLI to opt in; the default Pro context remains 200k. Max, Team, and Enterprise plans get 1M without the opt-in step. A developer evaluating whether Pro covers their workflow should plan for the opt-in step rather than assume 1M is on by default ⁸ . Enterprise contracts (Anthropic’s API and Team plans) follow separate negotiated pricing not covered in this article.

GST treatment for digital subscriptions billed by an Indian-resident user follows the standard 18% rate; the displayed Pro and Max sticker on claude.com may or may not be inclusive depending on the fulfilment path. Confirm at checkout. There is no India-specific student discount or regional pricing tier for either plan at the time of writing.

Image: Claude Plans and pricing page, used for editorial coverage of the Pro and Max tier framing.

What to test first

The temptation with a 5x context jump is to point it at a benchmark that measures recall-at-distance and call the result. That is the wrong test for an Indian engineering team deciding whether the upgrade pays back.

A better test sequence is three steps. First, pick a real multi-file refactor already on the team’s backlog, not a synthetic puzzle. Hand the entire repository to Claude Code in a single session and ask for a planning pass: “list the files that will need to change to do X, and explain why each one.” Compare the output against the file list a senior engineer would have hand-curated. Hits, misses, and false positives across that comparison are the load-bearing signal.

Second, run the same refactor through the 200k flow on the same codebase (Sonnet 4.5 or Opus at 200k) and time both. The wall-clock difference includes the model’s longer per-turn latency at 1M, the developer’s saved time from skipping pre-selection, and the rework cost from any context the model misses at either tier. The honest comparison is end-to-end, not per-turn.

Third, measure cost-per-completed-refactor in INR. The 1M context inflates per-turn input-token cost roughly proportionally; the question is whether the saved engineer hours exceed the higher inference bill. Indian engineering salaries are the cost denominator that makes this calculation different from the US-developer-centric framing in the vendor announcement. A senior engineer at ₹35-50 lakh CTC saves money for the company at a different breakeven point than the equivalent US developer at $200k.

Skip benchmarks like Needle-in-a-Haystack on synthetic prose. Real codebases stress long context in different ways than synthetic recall suites do, and the production wall-clock signal is what actually pays back.

Honest caveats and open questions

The 1M context number is taken from Anthropic’s public communications around the Opus 4.7 release; the precise documented ceiling, surface availability, and any per-tier limits could differ from the framing in this article and should be verified against the Anthropic news post and model documentation at the time of any purchase decision. The release shipped on 16 April 2026 and the picture will continue to sharpen as third-party reproducibility and benchmark coverage catches up over the following weeks.

The relationship between input context size and output behaviour at 1M is not yet well-characterised in independent reproducibility work. Anthropic publishes a model card for each release with internal benchmark scores; third-party evaluations on Indian-developer-relevant workloads (multi-language codebases, Hindi or regional-language documentation in repos, India-specific framework versions) are typically weeks to months behind the release. Expect early adopter reports to be noisy.

Claude Code’s CLI tooling is open source on GitHub ⁷ and the changelog will reflect any per-feature gating between Pro and Max tiers as it is published. Treat the GitHub repo’s documentation as authoritative for client-side behaviour; treat the Anthropic news post and model card as authoritative for server-side capability claims. Where the two diverge, the model card wins.

The Pro and Max delivered-cost figures in this article are computed from claude.com’s USD sticker prices plus the standard 18% GST and a typical 2-3% forex markup applied at the bank’s discretion. Indian payment-rail support (UPI, Razorpay rails, RuPay cards) is partial and inconsistently documented across vendor and third-party sources. The safe path is to attempt the purchase and abort if the rail you need is not supported, rather than relying on any third-party article, including this one, to confirm payment-method coverage on a moving target.