Claude Capability Tracker (Q2 2026)

Anthropic ships model updates faster than GTM teams adopt them. Most teams running Claude-based workflows in Q2 2026 are on capability that is one to two model generations behind. This tracker covers what shipped in Q1–Q2 2026 that is relevant to GTM practitioners — not a complete changelog, but a practitioner-filtered view of what to upgrade for and which model to use for which workflow.

The Q1–Q2 2026 Model Lineup

• Claude Opus 4.7 (1M context) — April 2026: Current flagship. Extended thinking GA, 1M-token context, improved reasoning over Opus 4.6, better tool-use reliability, and the new Agent Skills persistence framework. This is the model running this analysis.
• Claude Opus 4.6 — early 2026: Prior flagship. Still a strong reasoning model with extended thinking and 1M context. Baseline for complex agentic workflows; supports computer use in production.
• Claude Sonnet 4.6 — Q1 2026: The production workhorse. Cheaper than Opus, fast enough for real-time calls, strong at structured extraction, reply classification, and personalization generation. Default recommendation for production GTM pipelines.
• Claude Haiku 4.5 — October 2025 (claude-haiku-4-5-20251001): Cheap, fast tier. Pricing dropped sharply vs. Haiku 3.5. Best for high-volume classification, deduplication, simple enrichment. Performs within 5–10% of Sonnet 4.6 on clear-intent classification tasks at roughly 20% of the cost.

If you are still on any 3.x model, migration is almost always a one-line model string change. Stop reading and do that first.

Extended Thinking: Now GA Across the 4.x Family

Extended thinking — where the model reasons through a problem before generating output — is generally available across the 4.x lineup. It adds 4–15 seconds of latency and bills thinking tokens at output token rates. The question is not whether to use it but where.

Extended thinking earns its cost on: multi-signal account research synthesis (the company is expanding into EMEA but just announced a regional headcount reduction — what does that mean for propensity?), ambiguous multi-part reply intent classification, and complex ICP scoring with conflicting firmographic signals. For simple extraction — job title normalization, intent yes/no — it is unnecessary overhead. Opus 4.7 is the leading edge; the API flag is stable:

{
  "model": "claude-opus-4-7",
  "thinking": {
    "type": "enabled",
    "budget_tokens": 10000
  },
  "messages": [...]
}

A budget of 10,000 thinking tokens is sufficient for most complex GTM reasoning tasks. The thinking output is visible as a separate thinking content block in the API response — useful for debugging model decisions. At 16,000+ budget you are mostly adding cost with diminishing quality returns for standard GTM workflows.

Claude Opus 4.7 (1M Context): What the Context Window Unlocks

One million tokens is approximately: 12 months of Gong transcripts for a 10-person sales team, a 200-page RFP with all attachments, or a full multi-year account history including emails, notes, and CRM activity. The GTM use cases this unlocks:

• Full account history synthesis: Pass every interaction log, email thread, and call transcript for a strategic account in a single API call. No chunking, no retrieval-augmented lookup, no risk of missing a signal from 18 months ago.
• Multi-call recording synthesis: Pull all recorded sales calls for an account — discovery, demo, technical deep-dive, procurement — and generate a deal narrative, objection map, and stakeholder position summary in one call. This is a four-hour analyst task in 30 seconds.
• Full RFP analysis without chunking: Enterprise RFPs run 50–200 pages. Chunked retrieval loses cross-document context and misses requirements that span multiple sections. Opus 4.7 reads the full document and produces a compliance gap analysis and risk assessment in a single pass.

Cost math: at Opus 4.7 input pricing, a 500K-token account history synthesis run costs approximately $7.50–$15. For a strategic account where the deal is six figures, that is not a cost conversation worth having.

Agent Skills: Persistent State for GTM Workflows

Agent Skills is Anthropic’s persistent agent capability framework, released with Opus 4.7 in April 2026. The core change: Claude agents can carry state and tools across sessions rather than being stateless per-API-call systems. For AI SDR builds and multi-step GTM workflows, this changes the architecture in three ways:

• Multi-day account research: An agent can research a target account on Monday, check for new signals (funding announcement, job postings, earnings call) on Wednesday, and update its research brief on Friday — without rebuilding context from scratch. It remembers what it found and knows what changed.
• Adaptive lead nurture: A nurture agent can track what a prospect engaged with and what they ignored, adapting the next touch based on actual behavior rather than a fixed cadence template.
• Persistent tool bindings: Bind tools (CRM read/write, enrichment APIs, Gong) to an agent once and have them persist across sessions — no re-instrumenting on every invocation.

Agent Skills currently requires Opus 4.7 (or Opus 4.6 in some configurations). Budget for Opus pricing for multi-day persistent agents. Single-session agentic tasks remain Sonnet 4.6 territory.

Tool Use Reliability: The 4.x Step Change

Tool use reliability — correctly invoking the right tool with the right parameters on the first attempt — improved materially in the 4.x family. In production testing across Clay-equivalent enrichment workflows: Sonnet 4.6 reduces tool-call error rates by approximately 25–30% vs. Sonnet 3.7 and 40–45% vs. Sonnet 3.5. At 10,000 enrichment runs per month, that is 2,500–4,000 fewer failed runs requiring retry logic or human review.

JSON mode is GA across all 4.x models. Response schema enforcement (constraining output to a declared JSON schema) is robust. If you are still using prompt-hacked JSON extraction with regex parsers, remove them. Teams that migrated to native JSON mode reported 40–60% reduction in parsing failures:

{
  "model": "claude-sonnet-4-6",
  "response_format": { "type": "json_object" },
  "messages": [...]
}

Computer Use: Production-Ready in 4.6+

Computer use — Claude operating non-API SaaS UIs via screenshot-and-click interaction — is production-ready in Opus 4.6 and Opus 4.7. GTM-relevant use cases: automated data extraction from tools without cost-effective APIs, legacy CRM data entry, procurement portal navigation during RFP workflows, and enrichment lookups against tools that charge separately for API access. Computer use is slower than a native API integration. It is a deliberate fallback for when a native integration does not exist or is not cost-justified.

Practical Upgrade Checklist: Which Model for Which Workflow

• Account research / multi-document synthesis — Opus 4.7 (1M context): Full account history synthesis, RFP analysis, multi-call transcript synthesis, complex ICP scoring. Use extended thinking with 8,000–12,000 budget tokens. Enable Agent Skills for workflows spanning multiple days.
• Personalization at scale — Sonnet 4.6: Email personalization, LinkedIn message generation, call prep summaries, objection handling snippets. Enable JSON mode on all calls. At $3/$15 per million input/output tokens, Sonnet 4.6 is the price-performance sweet spot for personalization volume.
• Reply classification, deduplication, simple extraction — Haiku 4.5: Intent classification, lead dedup, job title normalization, company name standardization. Validate on 100 real examples from your dataset first — Haiku’s quality floor varies by task complexity. At volume, Haiku 4.5 should handle 60–70% of total API call volume.
• Agentic computer-use workflows — Opus 4.6+ with computer use enabled: Non-API SaaS navigation, legacy CRM workflows, procurement portals. Design human-in-the-loop checkpoints for any workflow writing data back to systems of record.

What Has Not Changed

Claude’s hallucination behavior on low-information inputs is not materially different in the 4.x family. Extended thinking improves reasoning on complex tasks; it does not grant the model access to information it does not have. Enrichment pipelines relying on Claude for factual research — rather than synthesis of data you provide — still require spot-checking and confidence thresholds. Latency on Opus 4.7 is 2–3x higher than Sonnet 4.6 on median full-response time. Match the model to the latency requirement, not just the capability requirement.

Forward Look: Q3 2026

• Agent Skills expansion: Broader availability across the model family, potentially including Sonnet 4.x, would make multi-day agent workflows cost-viable at Sonnet pricing.
• Computer use on Sonnet: Expansion from Opus-only would make agentic UI navigation practical for higher-volume workflows.
• Model refresh (likely): The 4.x cadence suggests a Sonnet 4.7 or Haiku 4.6 release in Q3. Teams on hardcoded model strings should plan for a migration touchpoint.
• Context caching improvements: Competitive pressure from Gemini and OpenAI will push further pricing optimization on long-context input. Teams running repeated long-context calls should watch for prompt caching updates that could materially reduce cost.

What to Ship This Quarter

Four actions, in priority order:

1. Audit model selection across all API calls. Map every workflow to the model it is using. Anything doing classification or simple extraction on Sonnet 4.6 should move to Haiku 4.5. This alone typically yields 30–50% API cost reduction with less than a day of engineering.
2. Enable JSON mode on all structured output calls. Remove your JSON parsing and retry logic. Fifteen minutes per workflow, eliminates an entire class of pipeline failures.
3. Pilot Opus 4.7 on one complex research workflow. Pick your highest-value account research use case, switch to Opus 4.7 with 1M context, and measure quality vs. your current chunked retrieval approach. The quality delta is usually obvious.
4. Evaluate Agent Skills for one multi-step workflow. Identify a workflow that rebuilds context from scratch on every run — account research refresh, multi-touch nurture — and prototype Agent Skills on it before Q3 planning locks.

The 4.x family is a meaningful generational step over 3.x. Most of the value is available through model selection optimization and native API feature adoption — not new architectural investment. Capture that before the next model cycle starts.

Related: See the full Anthropic/Claude vendor profile for current pricing tables, model comparison, and integration documentation. For a practical end-to-end build using these capabilities, see Building an AI SDR on Claude (Recipe).