If you already use Claude day-to-day and want it reasoning over your actual infrastructure — not a sandboxed demo, but your real AWS accounts, Kubernetes clusters, and Terraform modules — Clanker Cloud's BYOK setup is the direct path. You bring your Anthropic API key, choose claude-opus-4-6 or claude-sonnet-4-6, and Clanker Cloud handles the provider connections, context assembly, and query routing. Your credentials never leave your machine.
This article covers what changed in Claude 4.6, how to configure it in Clanker Cloud, and what you can actually do with it once it's connected.
Claude 4.6 — what actually changed
Claude Opus 4.6 shipped February 5, 2026. The headline benchmarks: 80.8% on SWE-bench Verified and 91.3% on GPQA Diamond. The more operationally significant number comes from METR's autonomous task evaluation: Opus 4.6 holds a 50%-time horizon of 14 hours and 30 minutes — the longest of any publicly available model at release. That figure matters for infrastructure work specifically. A model that maintains coherent reasoning across a 14-hour task window can hold an entire cloud environment in context without losing the thread on earlier findings.
The other significant addition is Agent Teams. Opus 4.6 can now spawn and coordinate multiple sub-agents working in parallel on a complex task. This is a native capability, not a third-party orchestration layer. For infrastructure audits, it maps directly to how the work should be decomposed: one sub-agent per provider, running simultaneously, results synthesized by the coordinating model.
Claude Sonnet 4.6 followed on February 17, 2026. The performance delta vs. Opus 4.6 on coding and document tasks is narrow. Where Sonnet 4.6 distinguishes itself is computer use: navigating browsers, filling forms, operating software GUIs. This is relevant for infrastructure tasks that require interacting with cloud consoles rather than just reading API data. Sonnet 4.6 also has meaningfully improved instruction following — it reliably does what you ask without interpretation drift across long multi-turn sessions.
API identifiers: claude-opus-4-6 and claude-sonnet-4-6.
Opus 4.6 vs. Sonnet 4.6 — choosing the right tier
| Scenario | Recommended model |
|---|---|
| Deep research audit across all providers | claude-opus-4-6 |
| IaC security review + remediation plan | claude-opus-4-6 |
| Six-month cost history analysis | claude-opus-4-6 |
| Day-to-day ops queries | claude-sonnet-4-6 |
| Real-time incident triage | claude-sonnet-4-6 |
| Cloud console navigation (computer use) | claude-sonnet-4-6 |
| Routine pod status / resource utilization | claude-sonnet-4-6 |
The decision is mostly about task horizon and cost. Opus 4.6 is the right tool for analysis that requires coordinating many findings across many resources. Sonnet 4.6 handles the rest at lower cost and faster inference.
Why Claude is a natural fit for infrastructure reasoning
The 200K context window is the starting point. An entire environment's worth of infrastructure configs, IaC files, cost exports, and recent logs can land in a single context. There is no chunking, no retrieval that loses the relationships between resources. Clanker Cloud assembles that context from your connected providers; Claude reasons over it.
Long-horizon task completion is the property that makes Claude 4.6 specifically valuable here, as opposed to earlier models. An infrastructure audit is not a single question. It is a sequence of observations, connections, and inferences that compound across dozens of resources. A model that drops earlier context produces findings that miss cross-resource relationships — the kind that produce the most consequential misconfigurations. Opus 4.6's 14hr 30min METR horizon means it does not.
Agent Teams maps directly to Deep Research. Clanker Cloud's Deep Research feature fans out across every connected provider and runs parallel analysis. Opus 4.6's native Agent Teams capability is the model-side complement to that architecture: the coordinating model spawns sub-agents per provider or per finding category, they run in parallel, and findings are aggregated and ranked by the orchestrator. The architectural fit is not coincidental.
Computer use in Sonnet 4.6 extends coverage beyond API-accessible data. Some infrastructure state is only visible through a cloud console UI. Sonnet 4.6 can navigate those interfaces directly — useful for tasks in providers or resource types where API coverage is incomplete.
Setting up Claude BYOK in Clanker Cloud
- Get your Anthropic API key at console.anthropic.com.
- Open Clanker Cloud and go to Settings → AI Model → Bring Your Own Key.
- Select Claude from the provider list.
- Paste your API key. It is stored locally and never transmitted to Clanker Cloud servers.
- Choose your default model tier:
claude-opus-4-6for deep research and complex audits,claude-sonnet-4-6for day-to-day queries.
You can switch between models per-session or set a default. The full configuration reference is in the docs.
Cost awareness. Opus 4.6 runs at approximately $15/1M input tokens and $75/1M output tokens (approximate 2026 pricing — verify current rates at console.anthropic.com). Sonnet 4.6 is substantially cheaper. A deep research scan pulling 50K–200K tokens of infrastructure context will cost roughly $0.75–$3.00 at Opus 4.6 input rates, plus output. For routine daily queries, Sonnet 4.6 brings that cost down by an order of magnitude. The recommendation in the FAQ: run deep research scans with Opus 4.6 on a scheduled or pre-event basis; use Sonnet 4.6 for everything else.
What Claude Opus 4.6 can do with your infrastructure
Deep Research scans
clanker ask "run a deep research scan across all my providers using Claude — find cost waste, misconfigs, and resilience gaps"
Opus 4.6's Agent Teams spawn parallel sub-agents per provider — one for AWS, one for GCP, one for Kubernetes, one for Cloudflare. Each sub-agent returns findings in its domain; the coordinating model ranks them by severity and cross-references dependencies. The output is a structured report with severity levels (medium, high, critical), affected resources, evidence sources, estimated cost impact, and concrete action labels. Findings export as Markdown or JSON for team sharing or ticket creation.
See the full capability set on the Deep Research use cases page.
Long-horizon cost analysis
clanker ask "analyze my entire AWS cost history for the past 6 months and identify structural waste patterns"
Six months of cost data across a non-trivial AWS account produces a large context. Opus 4.6 holds hundreds of resource configs simultaneously and maintains the thread between early observations and later findings. The result is not a list of idle EC2 instances — it is an analysis of which architectural decisions are producing recurring costs and what the structural remediation looks like.
IaC review and generation
clanker ask --maker "review my Terraform modules for security misconfigurations and generate a remediation plan"
Opus 4.6's 80.8% SWE-bench Verified score reflects genuine coding capability — it reads IaC as code, not as configuration prose. The --maker flag puts Clanker Cloud into plan mode: it assembles the remediation steps before executing anything. Run --apply to execute the plan once reviewed. For teams adopting vibe-coding-to-production workflows, this is the infrastructure side of that pattern.
What Claude Sonnet 4.6 can do with your infrastructure
Sonnet 4.6 handles the operational queries that happen continuously — the ones where latency matters and the task does not require multi-hour reasoning chains.
Day-to-day resource queries:
clanker ask "show me all EC2 instances with CPU utilization below 5% for the last 7 days"
clanker ask "what failed in my EKS cluster in the last hour"
Sonnet 4.6's faster inference means these come back quickly — important when you are in the middle of an incident and need answers without a multi-second model invocation.
Computer use. Sonnet 4.6 can navigate cloud consoles directly. For providers or resource types where the Clanker Cloud API integration does not cover a specific UI flow, Sonnet 4.6 can operate the console as a user would. This is useful for tasks like accessing preview features in GCP Console or navigating multi-step IAM policy editors in AWS.
Real-time incident triage. During incidents, the workflow is: query → observe → query again. Sonnet 4.6's inference speed supports that iteration cadence without the overhead of a deep-reasoning model. When an incident is resolved and a post-mortem analysis is needed, switch to Opus 4.6 for the longer-horizon work.
Teams building AI DevOps for teams workflows will typically configure Sonnet 4.6 as the default model and reserve Opus 4.6 for scheduled or on-demand deep research runs.
Claude via MCP for AI agent workflows
Clanker Cloud exposes an MCP endpoint that lets external AI agents — including Claude Code, Anthropic's coding agent — call infrastructure tools mid-session.
Start the MCP server:
clanker mcp --transport http --listen 127.0.0.1:39393
Once running, Claude Code can register Clanker Cloud as an MCP server. Infrastructure tools (clanker_route_question, clanker_run_command, clanker_version) become available inside the coding session without any context switching.
A concrete example: Claude Code is working through a deployment failure. It detects that a pod is not coming up. Without leaving the editor, it calls clanker_route_question with the relevant cluster and namespace context. Clanker Cloud returns live pod logs and recent events. Claude Code reads the logs, identifies the root cause, and applies the fix — all in a single session, no terminal switching, no manual console navigation.
Full documentation for agent integrations is at /for-ai-agents.md. The MCP setup is also covered in the Clanker Cloud docs.
Deep Research with Claude Opus 4.6 — the full flow
A concrete scenario: pre-launch infrastructure audit the week before a major release.
clanker ask "run a comprehensive deep research audit of my production infrastructure before next week's launch"
Opus 4.6's Agent Teams fan out: one sub-agent covers AWS services (EC2, RDS, S3, ALB), one covers Kubernetes workloads and resource limits, one covers Cloudflare DNS and WAF configuration. They run in parallel. The coordinating model aggregates findings and ranks them.
In a real scan against a production environment, the kind of findings this surfaces:
- Critical: RDS Postgres instance (
primary-db) has no automatic failover configured — single point of failure for the accounts database - High: EKS worker-pool resource limits too low for projected launch traffic — async job queue will saturate under load
- Medium: S3 bucket policy on
asset-bucketallows unintended cross-account read access
The team exports findings to Markdown, creates tickets from the critical and high findings, and resolves them before launch day. The entire analysis ran locally. The Anthropic API received the query text and the structured infrastructure data Clanker Cloud assembled — no raw credentials, no access keys, no secrets.
This is the workflow documented on the Deep Research use cases page, run with Claude Opus 4.6 as the reasoning layer.
Cost management for BYOK Claude
The practical split:
- Sonnet 4.6 for routine ops queries, incident triage, pod status, resource utilization checks
- Opus 4.6 for deep research scans, IaC audits, long-horizon cost analysis, and anything that requires Agent Teams parallelism
Clanker Cloud's Lite plan ($5/mo) and Pro plan ($20/mo) cover the workspace — provider connections, the CLI, the desktop app, MCP integration. Anthropic API costs are separate and accrue to your own API key. The Lite plan is sufficient for individual engineers running periodic deep research scans. The Pro plan is the right tier for teams sharing infrastructure context across multiple users.
A typical deep research scan — 50K–200K tokens of infrastructure context — costs approximately $0.75–$3.00 at Opus 4.6 input pricing, plus output tokens for the findings report. Running that scan weekly costs less than the Lite plan subscription. For routine daily queries with Sonnet 4.6, the cost is a rounding error.
More pricing detail and plan comparison in the FAQ.
FAQ
How do I use Claude Opus 4.6 with Clanker Cloud?
Go to Settings → AI Model → Bring Your Own Key in the Clanker Cloud desktop app. Select Claude, paste your Anthropic API key from console.anthropic.com, and set claude-opus-4-6 as your model. Your key is stored locally and never transmitted to Clanker Cloud servers. Once configured, all clanker ask queries route through Opus 4.6.
What is the difference between Claude Opus 4.6 and Sonnet 4.6 for infrastructure management?
Opus 4.6 is the right choice for tasks that require long-horizon reasoning and parallel sub-agent coordination — deep research audits, six-month cost analysis, IaC security reviews. It has a 14hr 30min METR task horizon and native Agent Teams support. Sonnet 4.6 is better for day-to-day queries, real-time incident triage, and tasks that benefit from computer use (navigating cloud consoles). Sonnet 4.6 is faster and significantly cheaper at inference time.
Does Clanker Cloud store my Anthropic API key?
No. The API key is stored locally on your machine and is never transmitted to Clanker Cloud servers. Clanker Cloud is a local-first application — provider credentials and AI API keys remain on the machine where the desktop app or CLI is running. Anthropic receives only the query text and the structured infrastructure data you ask it to analyze.
Can Claude's Agent Teams feature work with Clanker Cloud Deep Research?
Yes. When you run a deep research scan with claude-opus-4-6, Opus 4.6's Agent Teams capability maps directly to Clanker Cloud's parallel provider analysis architecture. The model spawns sub-agents per provider or per finding category — AWS, Kubernetes, Cloudflare, and others run simultaneously — and the coordinating model aggregates and ranks findings. The result is a severity-ranked report with affected resources, evidence, and remediation steps, exportable as Markdown or JSON.
Get started
Sign in or create an account at clankercloud.ai — the Clanker Cloud desktop app installs in under two minutes, and the CLI is available via Homebrew:
brew tap clankercloud/tap && brew install clanker
Connect your first provider, configure your Anthropic API key with BYOK, and run a deep research scan. The interactive demo at /demo shows the full flow if you want to see it before connecting live infrastructure.
For teams building AI-native infrastructure workflows, the full documentation is at docs.clankercloud.ai.
Give your agent live infrastructure context
Download Clanker Cloud, expose the local MCP surface, and let coding agents work from current cloud, Kubernetes, GitHub, and cost state instead of guesses.