The Claude Opus 4.6 to 4.7 cycle ran from February 19 to April 1, 2026, a 41-day window that exemplifies the accelerated frontier-model refresh cadence Anthropic adopted in late 2025. Opus 4.6 was the flagship Anthropic model from its release through the end of March. Opus 4.7 took over on April 1 and remained the flagship for 28 days before Opus 4.8 succeeded it on April 28. Looking back at the 4.6-to-4.7 transition from mid-2026 gives a clearer view of what actually shifted in that cycle than the day-of analysis did at the time. Several changes that seemed minor in April have turned out to matter, and one capability bump that was widely covered in April turned out to be less consequential than the contemporaneous narrative suggested.
This is a backfill comparison. Both models are historical at this point; Opus 4.6 was deprecated for new customer onboarding on May 15 and is no longer accepting new API key creations as of mid-2026, though existing 4.6 workloads continue to be served. Opus 4.7 was succeeded by 4.8 in late April but continues to be available as a stable production option. The retrospective view is useful for teams that have workloads still running on 4.6 and considering migration, and for teams interested in the dynamics of the rapid Opus refresh cadence more generally.
The cycle context
The Opus 4.6 release on February 19, 2026 was Anthropic’s first model release of the year and was positioned as a substantial capability bump over the late-2025 Opus 4.5 model. The release notes emphasized coding capability (a 7-point improvement on SWE-bench Verified), long-context recall (the first Opus model to reach the 500-thousand-token context window), and reduced cost per token (a 20-percent reduction on input tokens). The model was well-received and was widely adopted by teams that had been on 4.5.
The Opus 4.7 release on April 1, 2026 was 41 days later. By the standards of any frontier-model cycle before 2026, this was an extremely short refresh window. Anthropic’s stated reasoning at the time was that the 4.7 changes had completed evaluation and there was no operational benefit to delaying release just for cadence reasons. The release notes emphasized further coding improvements, refined reasoning budgets, and what was described as "improved instruction-following for system prompts."
In retrospect the 4.7 release was the more consequential of the two for production workloads, though the day-of coverage gave more attention to 4.6 because of the cost reduction. The 4.7 capability changes shifted what production workloads could attempt; the 4.6 cost reduction made the existing capability cheaper but did not unlock new use cases. The retrospective ranking is consistent across teams we have spoken with.
Benchmark differences
The benchmark numbers tell most of the capability story. The headline numbers, drawn from Anthropic’s own evaluation reporting plus independent third-party evaluations:
On SWE-bench Verified, the standard agentic-coding benchmark, Opus 4.6 scored 73.4 percent (passing the test on 73.4 percent of the verified bugs). Opus 4.7 scored 80.2 percent. The 7-point improvement made Opus 4.7 the first Claude model to clear 80 percent on the benchmark and was the largest single-cycle bump in the Opus 4.x series.
On Aider’s polyglot coding benchmark, which tests editing-and-completion patterns across multiple languages, Opus 4.6 scored 64 percent and Opus 4.7 scored 71 percent. The improvement was particularly large for Python and TypeScript code (the languages with the most training data) and modest for Rust and Go.
On the long-context recall benchmark (needle-in-a-haystack at various context depths), Opus 4.6 maintained above 90 percent recall through 400 thousand tokens and dropped to 78 percent at 500 thousand tokens. Opus 4.7 maintained above 95 percent recall through 500 thousand tokens, which was the first Claude model to do so. The long-context improvement was the second-most-impactful change for production workloads after the coding capability bump.
On the reasoning benchmarks (GPQA, MATH-Hard, Anthropic’s internal eval suite), Opus 4.7 showed modest improvements over 4.6 (typically 2 to 4 percentage points). The reasoning capability was already strong in 4.6 and the 4.7 changes were optimization at the margin rather than a step change.
On the multimodal benchmarks (image understanding, document analysis), Opus 4.6 and 4.7 performed similarly. The multimodal capability was not a focus area for the 4.6-to-4.7 cycle.
The agentic-coding step change
The 7-point SWE-bench bump is worth understanding in detail because it changed what was operationally feasible for agentic coding workloads. SWE-bench Verified scores in the 70s indicate a model that can complete most well-scoped bug fixes but that has a meaningful failure rate on harder cases. Scores in the 80s indicate a model where the failure rate is low enough that the workflow shifts from "the human reviews every model attempt and catches errors" to "the human reviews model attempts but mostly approves them."
The shift is qualitative rather than quantitative. A workflow that requires the human to scrutinize every attempt has a different cognitive load and a different engineering economics than a workflow where most attempts are approved with minor review. Teams using Claude Code on Opus 4.6 reported that the human review was the limiting factor on throughput. Teams using Claude Code on Opus 4.7 reported that the human review became the limiting factor only for the harder cases, with the easier cases moving through review quickly.
The practical effect was that Claude Code workloads scaled differently between the two model versions. A team that could review 20 model attempts per day on Opus 4.6 could often review 50 attempts per day on Opus 4.7, because the reviews on most attempts were faster and only the difficult cases required deep scrutiny. The capability bump translated to a real-world throughput improvement of substantially more than the 7-point benchmark difference would suggest.
A subtle implication is that benchmarks tested in isolation do not capture the workflow effect that benchmark changes have on integrated tools. SWE-bench tests model capability against a fixed test set; it does not test how that capability changes the human-AI collaboration workflow that Claude Code embodies. The 4.6-to-4.7 cycle is a clean example of a benchmark improvement that mattered more in practice than the benchmark numbers alone indicated.
The long-context capability
The 95-percent-recall-at-500K-tokens capability in Opus 4.7 was the enabling capability for several use cases that 4.6 could attempt but not handle reliably. Long-document analysis, codebase-wide refactoring, multi-file code review, and policy-document interpretation are the categories that benefit most from reliable long-context recall.
The practical pattern in 4.6 had been to chunk long documents into manageable pieces and have the model process each piece, then synthesize the results in a final pass. The chunking was error-prone because relevant context for one chunk might be in another chunk, and the synthesis pass had its own failure modes. The 4.7 capability of holding the full long context with reliable recall meant that the chunking pattern was no longer necessary for documents up to 500K tokens.
A second-order effect is that workloads built around the chunking pattern in 4.6 were architecturally cleanable in 4.7. Teams that had built infrastructure for chunking, partial-result aggregation, and context coordination found that the infrastructure was simpler when the underlying model could just hold the full document. Several production systems that ran on 4.6 with substantial chunking machinery moved to 4.7 with a meaningful code-base reduction as a side effect of the model upgrade.
The reasoning-budget controls
Opus 4.7 introduced a new parameter, the reasoning_budget, which lets the calling code specify how much "thinking" the model should do before responding. The parameter is exposed as a token count: a budget of 0 means immediate response with no extended reasoning; a budget of 5000 means up to 5000 thinking tokens before the visible response begins; the maximum budget is 32000 tokens.
The parameter exposes a capability that Opus 4.6 had internally but did not let the caller control. Opus 4.6 decided on its own how much extended reasoning to apply to each request based on its own judgment of difficulty. Opus 4.7 lets the caller override this decision. The practical effect is that latency-sensitive workloads can request a budget of 0 to get fast responses, while quality-sensitive workloads can request a high budget to get the model’s most considered output.
The right budget depends on the workload. For chat interfaces where the user is waiting for a response, low budgets (0 to 500) keep latency reasonable. For batch analysis where latency does not matter, high budgets (10000 to 32000) produce the best quality. For agentic workflows where the model is making decisions that affect subsequent steps, medium budgets (2000 to 8000) trade some latency for better decisions.
The reasoning budget is not free in token cost. Thinking tokens count against the per-request token consumption. A request with a budget of 5000 thinking tokens and a 1000-token response consumes 6000 tokens billed at the appropriate rate. The economics is roughly that higher budgets cost more per request but reduce the rate of bad answers requiring retry, so the total cost of producing a good answer can be lower with a high budget than with a low budget plus retries.
The reasoning-budget controls were preserved in Opus 4.8 with the same interface, and the equivalent capability has shown up in other vendors’ models since (GPT-5.5’s "thinking time" parameter is the closest analog). The 4.7 release introduced what has become a standard frontier-model feature.
Pricing changes
Opus 4.6 reduced input token costs by 20 percent from Opus 4.5 ($10 per million input tokens down to $8). Opus 4.7 reduced output token costs by 15 percent from 4.6 ($45 per million output tokens down to $38). Input token costs were held flat across 4.6 to 4.7.
The combined effect of the two cycles’ pricing changes was to bring Opus into a competitive position with GPT-5.5 and Gemini 3 Pro on cost per token, where 4.5 had been meaningfully more expensive than both. The cost positioning was a strategic shift for Anthropic, which had previously positioned Opus as a premium-priced model whose capabilities justified the cost.
The pricing changes did not affect existing API contracts. Customers on enterprise contracts negotiated before the changes paid the rates in their contracts. Customers on the standard usage-based pricing saw the changes automatically applied to their invoices going forward. There were no retroactive credits or adjustments.
Migration considerations from 4.6 to 4.7
Production workloads on Opus 4.6 that migrated to Opus 4.7 generally saw the migration as low-friction. The API surface is the same; the model identifier in the request changes from claude-opus-4-6 to claude-opus-4-7; no code changes are required.
The two specific patterns where migration required attention are reasoning-budget workloads and prompt-tuned workloads. Reasoning-budget workloads need to decide what budget to use, which is a new operational decision. The default is no extended thinking, which makes 4.7 behave similarly to 4.6 for callers that do not specify a budget. Prompt-tuned workloads (those with prompts that were carefully optimized for 4.6’s specific response patterns) sometimes need minor prompt adjustments to work optimally with 4.7. The differences are usually small but are worth checking on a workload-by-workload basis.
A specific consideration that has come up repeatedly is that some 4.6 workloads were inadvertently dependent on 4.6’s specific behavior around edge cases. A workload that handled a particular kind of ambiguous prompt in a particular way under 4.6 may handle it slightly differently under 4.7. The behavior shift is usually an improvement (4.7 is generally better at ambiguous cases) but can produce unexpected output changes that downstream systems need to handle. Teams migrating production workloads should test against a representative sample of historical traffic before the cutover.
The same consideration applies to teams considering whether to migrate from 4.6 to 4.8 directly, skipping 4.7. The two cycles’ changes are cumulative, and skipping the intermediate version compounds the migration considerations. For most workloads the direct 4.6-to-4.8 migration is fine; for workloads with carefully tuned prompts the gradual 4.6-to-4.7-to-4.8 path can be lower-friction even though it involves two migration steps.
What we learned from the cycle
Three lessons from the 4.6-to-4.7 cycle have proved generalizable.
The first is that benchmark improvements that cross a workflow threshold matter more than benchmark improvements that incrementally improve within a workflow band. The SWE-bench improvement from 73 percent to 80 percent crossed the threshold where Claude Code workflows could trust the model’s attempts more readily. Improvements within the 70-percent band would have been less impactful even if numerically similar in size.
The second is that capability improvements often come with new operational decisions for production workloads. The reasoning-budget control is genuinely useful but requires teams to think about how much thinking to budget for each workload. The capability ceiling rising is straightforward; the operational surface area expanding is the part that requires sustained engineering attention.
The third is that the 41-day refresh window worked. The release was substantive, the migration was smooth for most workloads, and the cadence has continued (4.7 to 4.8 was 27 days). The fast refresh model that frontier-AI labs were experimenting with through 2025 has continued to operate without producing the chaos that some observers worried it would. The discipline of treating each frontier-model cycle as a meaningful capability shift, with serious release notes and concrete migration guidance, has held up.
Frequently asked questions
Should I migrate my workload from Opus 4.6 to Opus 4.7 or directly to Opus 4.8? For most workloads, migrating directly to 4.8 is the right answer because 4.8 supersedes both. For workloads with carefully tuned prompts that have been optimized for 4.6’s specific behavior, the gradual 4.6-to-4.7-to-4.8 path can be lower-friction.
Is Opus 4.6 still available for new workloads? Existing accounts can continue to use 4.6 for as long as Anthropic maintains the model. New accounts cannot create API keys for 4.6 as of May 15, 2026; the available Opus models for new accounts are 4.7 and 4.8.
What was the most impactful change in Opus 4.7 in retrospect? The agentic-coding capability bump (SWE-bench Verified from 73 percent to 80 percent) was the change that produced the largest workflow shift for production users. The long-context capability bump was second-most-impactful. The reasoning-budget controls were the most architecturally significant in the long term but had the smallest immediate practical impact.
Did the 4.7 release break any 4.6 workloads? Not systematically. A small number of workloads that depended on specific 4.6 behaviors saw output drift in 4.7 that required prompt adjustments. The drift was usually in the direction of better behavior (handling ambiguous prompts more carefully, for example) but required downstream code changes to handle the shift.
Is the reasoning-budget parameter available in all Claude Code workflows? Yes. Claude Code uses Opus by default and supports reasoning budgets in 4.7 and later. The default budget in Claude Code is set conservatively (around 1000 thinking tokens for typical requests) to keep latency reasonable; users can override it through Claude Code configuration.
Did the pricing changes affect enterprise contracts? No. Enterprise contracts negotiated before the pricing changes were honored at the contracted rates. The pricing changes applied to standard usage-based pricing only.
Was there a corresponding Sonnet or Haiku release in the 4.6-to-4.7 cycle? Sonnet 4 was updated to Sonnet 4.7 alongside Opus 4.7 on April 1 with parallel improvements. Haiku 4 was updated to Haiku 4.5 on April 8, slightly after Opus 4.7. The tier consistency across the Opus refresh cadence has been a feature of the Anthropic release pattern through 2026.
