AI Product Updates, July 25: Claude Opus 5 Lands Across the Stack (2026)

The short version

Claude Opus 5 is the main model release in this update cycle, and it is already spreading beyond Anthropic's own products: GitHub Copilot is adding it to its model picker and Amazon Bedrock lists it as an active model. Google also moved Apigee API hub's MCP server to general availability, while Replit added Slack-aware Agent workflows and set lower Cloud prices for billing cycles that begin on or after August 1. 1 2 3 4 5

Product or platform	Releasing entity	Release date	Core capability or change	Availability status
Claude Opus 5	Anthropic	July 24, 2026	New Opus-tier model for long-running agents, coding, and professional work; $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.	Available in Claude, the API, Claude Code, and Claude Cowork. It is the default model for Max and the strongest model on Pro. 1
Claude Opus 5 in Copilot	GitHub	July 24, 2026	Adds Opus 5 to the model picker for complex, multi-step coding tasks, with enhanced safeguards for high-harm cyber content.	Rolling out to Copilot Pro+, Max, Business, and Enterprise; Business and Enterprise admins must enable the policy. 2
Claude Opus 5 in Bedrock	Amazon Web Services	July 24, 2026	Active model ID `anthropic.claude-opus-5`; Bedrock's model card lists a 1M-token context window, 128K maximum output, adaptive thinking, and prompt caching.	Available through Bedrock's supported runtime and Mantle routes, subject to AWS account and regional access. 3
API hub MCP server	Google Cloud Apigee	July 24, 2026	MCP tools can discover, query, create, update, and delete APIs, versions, specs, and deployments; a new RPC can configure and deploy MCP servers to Apigee.	GA through `apihub.googleapis.com/mcp`, with read-only and read-write OAuth scopes and Model Armor integration. 4
Replit Agent and Cloud	Replit	July 24, 2026	Mobile Agent adds voice-driven building and better progress notifications; the Slack integration can search accessible context and act on the user's behalf. Cloud usage prices drop from August 1.	Mobile and Slack features are available through their respective integrations. New Cloud rates apply automatically at the first billing cycle beginning on or after August 1, across all plans. 5 6

Claude Opus 5 is becoming a deployment choice, not just a model choice

Anthropic says Claude Opus 5 is available in Claude, the Claude API, Claude Code, and Claude Cowork. The launch is aimed at long-running agent work, software engineering, debugging, multi-step tasks, and professional knowledge work. The listed API price is $5 per million input tokens and $25 per million output tokens, the same price as Opus 4.8. Fast mode runs at about 2.5 times the default speed and costs twice the base price. 1

Two API behavior changes matter as much as the model itself. Mid-conversation tool changes can now happen without invalidating the prompt cache. Automatic fallbacks can route requests that a safety classifier flags to another model instead of returning a hard block. Claude.ai, Claude Code, and Claude Cowork default to Opus 4.8 for those fallbacks; API users can enable the behavior themselves. 1

That fallback path changes how applications should log failures. A request that used to stop may now complete on a different model, so production systems should record the selected model and evaluate outputs from both the primary and fallback paths.

GitHub puts Opus 5 inside the coding loop

GitHub is rolling Claude Opus 5 into Copilot Pro+, Max, Business, and Enterprise. The model is selectable across Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. The rollout is gradual, and Business and Enterprise administrators must enable the Claude Opus 5 policy before users can select it. 2

GitHub describes the target workload as complex, long-running coding tasks that require tool use, autonomous changes, regression verification, and repeated validation. It also warns that the model's high-harm cyber safeguards may block some security-adjacent requests. Copilot bills Opus 5 at the provider's API list price under usage-based billing, so teams should check the current model-pricing page rather than assume it is included in a flat plan allowance. 2

The practical test is straightforward: enable it for a small repository, compare autonomous change quality and regression verification against the model your team uses now, and track request costs separately from seat fees.

Bedrock exposes a different Opus 5 contract

Amazon Bedrock lists Claude Opus 5 as an active model with the ID anthropic.claude-opus-5. Its model card lists a 1M-token context window, 128K maximum output, adaptive thinking enabled by default, computer use, prompt caching, and launch support on July 24. Prompt-cache checkpoints start at 512 tokens, with five-minute and one-hour TTL options. 3

The interface details are the part to check before migrating. Bedrock Runtime supports Invoke, Converse, Messages, and Text for this model, but not Responses or Chat Completions. The model card also marks structured outputs and intelligent prompt routing as unsupported on Bedrock Runtime. Mantle supports streaming, abuse detection, token counting, and computer use, but has a different set of exclusions. 3

AWS documents US, EU, Australia, and global routing options, with access still subject to an AWS account and supported region. The Bedrock page does not state a numeric price; use the live Bedrock pricing page for a cost comparison rather than carrying Anthropic's direct API price into an AWS estimate. 3

Apigee makes MCP an API-management surface

Google Cloud's Apigee API hub MCP server is now GA. Agents can use natural language to discover, query, and manage an API ecosystem, including creating, updating, and deleting APIs, versions, specs, and deployments. The release also adds ConfigureAndDeployServer, which can configure and deploy MCP servers directly to an Apigee runtime. 4

The access model is more specific than a generic MCP connector. The global endpoint is apihub.googleapis.com/mcp, the release adds apihub.readonly and apihub.readwrite OAuth scopes, and Model Armor can protect MCP tool invocations from prompt injection and related attacks. Teams should start with the read-only scope, test discovery and query flows, and grant write access only to an agent with an explicit deployment approval path. 4

Replit connects Agent to Slack and changes the cloud bill

Replit's July 24 changelog adds a refreshed mobile home screen, simpler navigation between Agent, tasks, and Preview, voice prompts for building, and improved notifications with Live Activities. Its Slack integration gives Agent access to the messages, files, channels, and people the connected user can access. Agent can read private conversations and thread replies, send messages as the user, and create or update Slack canvases. 5

The permission boundary deserves a real test: this is an action-capable integration, not just a Slack search box. A pilot should use a workspace with limited channels and a test identity before allowing Agent to send messages or edit canvases.

Replit also published the old and new Cloud rates. The reductions take effect automatically at the start of the first billing cycle on or after August 1, 2026, and apply to Core, Pro, Teams, and Enterprise plans. A few representative changes are:

Replit Cloud charge	Current price	New price from August 1
Autoscale compute units, per million	$3.20	$0.60
Autoscale requests, per million	$1.20	$0.40
Outbound data transfer, per GiB	$0.10	$0.05
App Storage data, per GiB-month	$0.03	$0.015
Database compute, per compute hour	$0.16	$0.16

The database rates are unchanged; the other figures above come from Replit's current-versus-new comparison table. Published projects switch automatically, but the exact transition date follows each account's billing-cycle start. 6

What to test next

Anthropic API: test a tool change mid-conversation and log whether prompt-cache behavior is preserved. If automatic fallbacks are enabled, record the actual serving model for every request.
GitHub Copilot: have an administrator enable Opus 5 for a small repository, then compare tool calls, regression checks, cyber-related blocks, and usage cost with the current default.
Amazon Bedrock: audit the client path before migration. Code written for Responses or Chat Completions will need a supported Bedrock interface instead.
Apigee: begin with apihub.readonly and Model Armor. Treat create, update, delete, and deployment tools as separate approval scopes.
Replit: test Slack with a restricted workspace and wait for the first post-August billing cycle before updating a production cost forecast.

The short-term pattern is clear: Opus 5 is available in more places, but each host exposes a different price, interface, safety policy, or rollout gate. Treat the model name as the start of an integration check, not the end of one.

AI Product Updates, July 25: Claude Opus 5 Lands Across the Stack