Claude 4 and 4.5: Computer Use, MCP, and the Anthropic-Amazon Bet

Anthropic shipped two consequential releases in the back half of 2025 that reset where Claude sits in the enterprise AI stack. Claude Sonnet 4.5 landed on September 29, 2025 with stronger coding performance and meaningful agentic improvements. Claude Opus 4.5 followed in November with a one-million-token context window and the strongest long-horizon reasoning of any frontier model shipping today. Surrounding both releases sit three things that arguably matter more than the model versions themselves: the Computer Use capability that turned Claude into a genuine agentic surface, the Model Context Protocol that Anthropic open-sourced and the industry quietly adopted, and the multi-billion-dollar deepening of the Anthropic-Amazon partnership.

This is what changed and what it means for teams building production AI in 2026.

Sonnet 4.5 and Opus 4.5 in practice#

Sonnet 4.5 was Anthropic’s coding-and-agent release. Anthropic published a 77.2% score on SWE-bench Verified at launch with parallel test-time compute, ahead of GPT-5 Codex on the same benchmark. The internal-tool-use and computer-use benchmarks also stepped up meaningfully. Pricing held at the same three-dollars-per-million-input and fifteen-dollars-per-million-output rate that Sonnet 4 launched at, which keeps it as one of the better cost-quality points in the market.

Opus 4.5 in November pushed the frontier in a different direction. The context window expanded to one million tokens — matching Gemini 2.5 Pro and exceeding GPT-5 Pro’s 400K — and the model’s behavior on long-horizon agentic tasks improved enough that several enterprise customers shifted from multi-step orchestrated pipelines to single-call Opus invocations for end-to-end work. The benchmark numbers were strong; the practical difference for production teams was the stability of multi-hour autonomous sessions where the agent maintains coherence across hundreds of tool calls.

The tier structure that settled by year-end:

• Claude Haiku 4.5 — fast, cheap, used for high-volume classification and extraction
• Claude Sonnet 4.5 — the workhorse default, used for the bulk of agentic and reasoning workloads
• Claude Opus 4.5 — the deepest-reasoning tier with 1M context, used for research-grade work and long-running agents

Computer Use and the agentic surface#

Anthropic launched the Computer Use capability publicly in October 2024 and matured it through 2025 alongside the 4.5 releases. The capability lets Claude take screenshots of a virtual machine, move the mouse, type into fields, and navigate browser and desktop applications. It is the first frontier-model capability that meaningfully crosses from text generation into direct OS-level action.

Adoption in 2025 was experimental rather than mainstream. The latency is meaningful, the reliability on novel UIs is uneven, and the security posture of giving a frontier model direct desktop control is something most enterprise security teams have not yet figured out how to bless. But the capability is real, the pattern is reproducible, and several SaaS vendors have shipped Computer-Use-powered features that automate workflows their customers could not previously automate at all. The agentic-UI space is one of the most-watched product surfaces going into 2026.

The Model Context Protocol#

Anthropic open-sourced the Model Context Protocol in November 2024. By the second half of 2025 it had quietly become the de facto standard for connecting models to tools and data sources. OpenAI announced MCP support for ChatGPT and the Agents SDK in March 2025, Google DeepMind confirmed MCP support for Gemini in April 2025, and Microsoft added native MCP support across Copilot Studio, GitHub Copilot, and Windows in mid-2025. The protocol’s ascendance is one of the more unusual stories of the year — an Anthropic specification adopted by its largest competitors with little drama.

For developer-tool builders the practical effect is real. Building an integration once against MCP now exposes it to Claude, GPT, Gemini, and Copilot. The cost of supporting multiple model providers has dropped sharply for the integration layer.

The Anthropic-Amazon partnership#

The financial and strategic story underneath all of this is the deepening of the Anthropic-Amazon relationship. Amazon committed an additional 4 billion dollars to Anthropic in November 2024 on top of the earlier 4 billion, taking the total commitment past 8 billion dollars. The structural condition was that Anthropic shift the bulk of its training and inference workloads onto AWS — specifically onto Trainium and Inferentia silicon — and that AWS become the primary cloud partner for the Anthropic-trained models.

The training infrastructure for Claude is now running on AWS Trainium2 clusters in the Project Rainier supercluster — Anthropic confirmed this publicly through 2025. The implication for the broader market is that AWS now has a frontier-model partner the way Microsoft has OpenAI and Google has its own Gemini stack. Bedrock customers get the most current Claude models without the procurement overhead of a separate Anthropic contract, and the integration cost-of-switching from OpenAI on Azure to Claude on Bedrock has dropped meaningfully.

Developer tool adoption#

The single strongest signal of where Claude sits in 2026 is developer-tool adoption. Cursor, GitHub Copilot’s agentic mode, Anthropic’s own Claude Code CLI, Replit Agent, Vercel v0, and a long tail of smaller coding tools have made Claude Sonnet 4.5 their default for non-trivial work. The reasons compound: the model is genuinely strong on SWE-bench, the long context handles full-repo loads, the tool-use API is polished, and the pricing point lets product teams build agentic features without burning their cost envelope.

For enterprises this matters because the developer-experience surface is where AI adoption either spreads or stalls. Engineering teams using Claude for daily coding work pull Claude into the broader product engineering decisions; teams who hated their first AI coding tool refuse to revisit the question for years.

Where pdpspectra fits#

Our AI and LLM integration practice ships production Claude deployments across healthcare, banking, and SaaS clients — typically with Sonnet 4.5 as the default workhorse, Haiku 4.5 for high-volume operations, and Opus 4.5 reserved for the hardest reasoning workloads. We use Claude via AWS Bedrock for regulated AWS-native clients and via the direct Anthropic API for clients that want the fastest model rollout and the cleanest workspace structure.

Related reading: Bedrock vs OpenAI vs Anthropic for enterprise, GPT-5 implications, and the open-source LLM landscape.

Closing#

Claude in 2026 is no longer the underdog frontier model. The 4.5 releases, the MCP protocol adoption, the Computer Use capability, and the Amazon partnership have made it one of the two or three default choices for serious enterprise AI work. The developer-tool surface has tilted decisively in Anthropic’s direction, and Bedrock has made the procurement path frictionless for AWS-native enterprises.

For teams choosing a frontier model in 2026, Claude deserves a real evaluation rather than the second-place benchmark it had in 2023. Talk to our team about your model selection.