Anthropic's spring 2026 lineup: Opus 4.7, Sonnet 4.6, Haiku 4.5, and where each model slots in

The bottom line

Anthropic’s spring 2026 Claude lineup has three tiers and a clear price-per-tier story. Opus 4.7 is the top tier, generally available since 16 April 2026, priced at $5 per million input tokens and $25 per million output tokens; the improvements over Opus 4.6 are concentrated in software engineering, long-running tasks, and higher-resolution vision. ¹ Sonnet 4.6 is the balanced tier, released 17 February 2026 at $3 input / $15 output per million tokens, with a 1M-token context window in beta; Anthropic’s own framing says developers preferred it to Sonnet 4.5 by roughly 70% and over Opus 4.5 by about 59% for many tasks. ² Haiku 4.5 remains the cheap-and-fast tier, released 15 October 2025 at $1 input / $5 output per million tokens. ³

For anyone reading this hoping for a separately-numbered “Claude Haiku 4” launching in May 2026: there is no such model. Anthropic skipped a 4.0 numbering for Haiku and went straight to 4.5, which is still the current Haiku tier as of 19 May 2026. ^{3 4}

The actual May 2026 news cadence is everything else around the model lineup. Code w/ Claude SF on 12 May doubled Claude Code rate limits and shipped four new Managed Agents capabilities. Fast mode in the Developer Platform now uses Opus 4.7. Claude Platform on AWS hit general availability on 11 May. Microsoft 365 integration went GA across Excel, PowerPoint, and Word. ⁵

Prices verified at vendor pricing pages as of 19 May 2026; prices fluctuate, verify before purchase.

Where each model slots in

Opus 4.7 is positioned for the deepest reasoning workloads: codebase refactoring, multi-agent coordination, long-running coding tasks. Anthropic’s launch post claims a 13% improvement on its 93-task coding benchmark over Opus 4.6 and notes four tasks “neither model nor competitors could previously solve.” Vision capability rose to processing images up to 2,576 pixels on the long edge — roughly 3.75 megapixels — which the company frames as “more than triple previous versions.” ¹

Sonnet 4.6 is positioned as the workhorse tier: a “full upgrade of the model’s skills across coding, computer use, long-context reasoning, agent planning, knowledge work, and design,” per Anthropic’s launch post. The 1M-token context window in beta is the headline new affordance. Sonnet 4.6 keeps the same $3 / $15 per-million-token pricing as Sonnet 4.5 — no price increase for the upgrade. ²

Haiku 4.5 is positioned as the speed-and-cost tier. Anthropic’s launch positioning calls it “near-frontier performance with much greater cost-efficiency” and notes Haiku 4.5 runs “up to 4–5 times faster than Sonnet 4.5 at a fraction of the cost.” On SWE-bench Verified, Haiku 4.5 scores 73.3% in Anthropic’s reported numbers, which puts it at the top end of small-model coding performance. ³

Pricing table

| Model | Released | Input ($/MTok) \mid Output ($/MTok) | Headline positioning | |---|---|---|---|---| | Opus 4.7 | 16 April 2026 | $5 \mid $25 | Deepest reasoning; long-running coding; large-context refactoring | | Sonnet 4.6 | 17 February 2026 | $3 \mid $15 | Balanced workhorse; 1M context (beta); preferred over 4.5 by ~70% | | Haiku 4.5 | 15 October 2025 | $1 \mid $5 | Cheap-and-fast; ~4–5x faster than Sonnet 4.5; 73.3% on SWE-bench Verified |

Pricing as of 19 May 2026, from each model’s launch post and the Claude Platform models overview. ^{1 2 3 4} Prices fluctuate, verify before purchase.

The actual May 2026 news cadence

The model-tier story stayed put through May, but the surrounding product surface moved a lot. Per Anthropic’s release notes aggregated by Releasebot: ⁵

11 May 2026 — Agent View in Claude Code. A new CLI feature letting users manage multiple concurrent Claude Code sessions from a single dashboard. Background execution, status peeks, and selective resume without juggling terminal tabs.

11 May 2026 — Claude Platform on AWS GA. The Claude API became available through AWS authentication and billing, with native parity for Managed Agents, code execution, skills, and prompt caching.

12 May 2026 — Code w/ Claude SF Recap. Anthropic doubled Claude Code rate limits and raised API limits for Opus. New Managed Agents capabilities: Dreaming (reviewing past sessions for patterns), multi-agent orchestration (lead agents delegating to specialist sub-agents), Outcomes (rubric-based output grading), and webhooks for notifications.

12 May 2026 — Fast mode on Opus 4.7. The Developer Platform’s Fast mode research preview switched from Opus 4.6 to Opus 4.7 as the underlying model. Same pricing tier, faster output generation.

12 May 2026 — Claude for Legal Industry. 20+ MCP connectors covering contract lifecycle, e-discovery, document management, and legal-research platforms. 12 practice-area plugins. Microsoft 365 integration for Word, Outlook, Excel, and PowerPoint.

13 May 2026 — Claude for Small Business + cache diagnostics. Small-business package with integrations across QuickBooks, PayPal, HubSpot, and Microsoft 365; 15 prebuilt workflows. Separately, prompt-cache diagnostics in public beta on the Developer Platform with cache_miss_reason explanations.

18 May 2026 — Richer SEC filing data in web search. The web search tool now surfaces deeper SEC filing structure for financial analysis workflows.

19 May 2026 — MCP tunnels in research preview. The Developer Platform launched MCP tunnels (research preview) along with self-hosted sandboxes for Managed Agents with live MCP configuration updates.

What the May cadence says about Anthropic’s strategy

Two themes stand out across the May 2026 announcements.

First, Managed Agents is now a first-class product, not a research preview. Doubled rate limits, four new agent capabilities, Agent View dashboard, multi-agent orchestration, webhooks. Anthropic is treating long-running multi-step Claude Code sessions as the primary developer workflow and shipping the orchestration plumbing to match.

Second, MCP is the integration layer. The Legal Industry release ships 20+ MCP connectors. The Small Business release uses MCP-style integration patterns. The Developer Platform’s 19 May announcement introduces “MCP tunnels” as a transport. Almost every May release that involves a third-party data source is built on MCP rather than on a one-off integration.

Neither theme depends on a new model release. Opus 4.7 carries the heavy reasoning work and Sonnet 4.6 carries the production volume; the May cadence is about turning those models into deployable agent systems.

Comparison to the prior cadence

Model	Prior version (when)	Headline change	Price change
Opus 4.7 (16 Apr 2026)	Opus 4.6	+13% on Anthropic’s coding benchmark; higher-res vision; better long-running task reliability	None
Sonnet 4.6 (17 Feb 2026)	Sonnet 4.5	1M context (beta); coding, computer-use, agent-planning upgrades; preferred over 4.5 ~70% of the time	None
Haiku 4.5 (15 Oct 2025)	Claude 3.5 Haiku	Matches Sonnet 4 on coding; 4–5x faster than Sonnet 4.5	Increase vs 3.5 Haiku (which was $0.80 / $4 per MTok); see Claude Platform pricing for full history

Sources for each row: Opus 4.7 launch post, Sonnet 4.6 launch post, Haiku 4.5 launch post. ^{1 2 3} Verify pricing history on the Claude Platform models overview before any planning decision. ⁴

How to read the lineup if you are choosing a model today

Sources flag three default heuristics that the Anthropic documentation supports: ⁴

For deep reasoning, long context, multi-agent orchestration, Opus 4.7 is the tier Anthropic positions for the task. The $5 / $25 per-MTok price is the cost of that depth.

For production workflows with mixed reasoning and tool use, Sonnet 4.6 is the tier. Anthropic’s own preference numbers (Sonnet 4.6 chosen over Opus 4.5 about 59% of the time per its launch post) suggest Sonnet 4.6 is the default unless the workload genuinely needs Opus. ²

For high-volume, latency-sensitive, cost-sensitive workloads, Haiku 4.5 remains the tier and has not been superseded. Its $1 / $5 pricing and 4-5x speed advantage over Sonnet 4.5 are the load-bearing reasons. ³

Honest caveats

Anthropic’s own benchmark numbers (13% improvement, 70% preference, 73.3% on SWE-bench Verified) are the publisher’s reported results. Independent reproductions exist for SWE-bench Verified through the public leaderboard but should be verified separately for any specific workload. The “preferred 70% of the time” framing in particular is internal A/B-test framing that is not directly verifiable from public artefacts.

This article covers Anthropic’s spring 2026 lineup as of 19 May 2026. The cadence may produce a Sonnet 4.7, a Haiku 5.0, or any other variant after this date; the Claude Help Center release-notes page is the canonical source for new model announcements. ⁶

The May 2026 product announcements (Code w/ Claude SF recap, Claude for Small Business, MCP tunnels research preview) ship continuously; specific features may move between research preview and GA. Verify on the Developer Platform release-notes page before depending on any specific capability in production.