Skip to content
Uvilo OS
published Visibility internal Owner erik@uvilo.com Approver _ Created _ Updated _

10. The Model Landscape — What to Use, When, and Why It Keeps Changing

Last verified: April 8, 2026. Model pricing and availability shift fast. If you’re reading this more than 60 days after that date, verify the numbers yourself before building around them.

The Lie You’ll Hear First

“Just use [Model X] for everything.”

You’ll hear this from Twitter threads, YouTube thumbnails, and AI newsletters. It’s wrong. Not because Model X is bad, but because “everything” is the problem.

An operator who sends morning summaries, runs parallel coding agents, replies to customers on Slack, triages Sentry alerts, and writes marketing copy doesn’t have one job. They have five jobs with completely different performance, cost, and latency requirements. Routing all of them through one model is like hiring a surgeon to also answer phones, stock shelves, and mop floors. The surgeon is talented. That doesn’t make it efficient.

The first serious thing you learn as an operator: route by job, not by loyalty.

The Models That Actually Matter Right Now

These are the frontier models available via API as of April 2026, with real pricing pulled from official docs. Not vibes, not benchmarks someone screenshotted — the numbers you’ll see on your invoice.

Anthropic Claude

ModelInput / Output (per MTok)ContextMax OutputBest For
Claude Opus 4.6$5 / $251M tokens128K tokensComplex agentic work, deep reasoning, coding architecture
Claude Sonnet 4.6$3 / $151M tokens64K tokensBest speed/intelligence balance, daily operator work
Claude Haiku 4.5$1 / $5200K tokens64K tokensFast cheap work, summaries, classification

Opus 4.6 is Anthropic’s strongest broadly available model. Knowledge cutoff May 2025, training data through August 2025. Sonnet 4.6’s training data goes through January 2026 — it actually knows more recent events than Opus.

Both Opus 4.6 and Sonnet 4.6 support extended thinking (explicit chain-of-thought reasoning budgets) and adaptive thinking (the model decides how much to think based on the problem). Haiku 4.5 supports extended thinking but not adaptive.

All three support text + image input. All three are available on direct API, AWS Bedrock, and Google Vertex AI.

Legacy still available: Sonnet 4.5, Opus 4.5, Opus 4.1, Sonnet 4, Opus 4. But Haiku 3 is deprecated. If you’re still on Haiku 3, migrate.

The operator note: Claude’s 1M context window on Opus and Sonnet is real and usable — not a marketing number where quality degrades at 200K. This matters for codebases with many files and long conversation histories.

OpenAI GPT

ModelInput / Output (per MTok)ContextBest For
GPT-5.4$2.50 / $15 (short) · $5 / $22.50 (long)—Complex reasoning, coding, flagship tasks
GPT-5.4 mini$0.75 / $4.50400K tokensHigh-throughput work at lower cost
GPT-5.4 nano$0.20 / $1.25—Edge/embedded, cheapest possible
GPT-5.4 pro$30 / $180—Maximum capability, premium pricing

GPT-5.4 released March 5, 2026. OpenAI’s pricing page separates “short context” and “long context” tiers for GPT-5.4 — once your prompt exceeds a threshold, input and output pricing jump. Plan for this if you run long sessions.

Codex lane: OpenClaw’s provider docs treat openai-codex/gpt-5.4 as the current Codex subscription path. That’s worth thinking about as a separate coding lane from direct openai/* API traffic, because the auth path, economics, and session behavior differ.

Cached input pricing: GPT-5.4 drops to $0.25 per MTok for cached input (90% savings). GPT-5.4-mini caches at $0.075. If your workflow involves repeating large system prompts, this adds up fast.

Google Gemini

ModelInput / Output (per MTok)ContextBest For
Gemini 3.1 Pro (preview)$2 / $12 (≤200K) · $4 / $18 (>200K)1M tokensStrongest Google reasoning, multimodal
Gemini 3 Flash (preview)$0.50 / $3—Balanced speed and capability
Gemini 2.5 Pro$1.25 / $10 (≤200K) · $2.50 / $15 (>200K)1M tokensProven, stable, good value
Gemini 2.5 Flash$0.30 / $2.501M tokensFast daily work, great cost ratio
Gemini 2.5 Flash-Lite$0.10 / $0.40—Cheapest serious model on the market

Gemini 3.1 Pro launched February 19, 2026. Google claims 2x+ reasoning boost over Gemini 3 Pro and ranks #1 on 12 of 18 tracked benchmarks. It supports 1M context with 65K token output.

The Google advantage that gets overlooked: Gemini is natively multimodal in a way the others aren’t. Text, images, video, audio, code — all in the same model. If your agent workflow involves screenshots, image generation, video understanding, or grounded web search, Google should be your media lane even if it’s not your primary text lane.

Free tier: Google AI Studio offers free access to Gemini 2.5 Flash, 2.5 Flash-Lite, and 3.1 Flash-Lite with rate limits. This is legitimately useful for testing and low-volume side projects.

Context caching: Google absolutely offers context caching, but the official docs price it by model, token volume, and storage duration. In other words, it is real, but it is not a universal “90% off” rule. Budget it from the pricing table instead of repeating one magic percentage.

xAI Grok

ModelInput / Output (per MTok)Best For
Grok 4.20$2 / $6 (cached input $0.20)xAI’s current flagship reasoning lane
Grok 4.1 Fast$0.20 / $0.50 (cached input $0.05)Cheap overflow work, fast second opinions

xAI’s own docs currently call Grok 4.20 the newest flagship model. Grok 4.1 Fast is the cheap tier. Those are not the same price class, and collapsing them into one line badly distorts budgeting.

The catch: xAI’s pricing docs are clear, but its public benchmark story is still thinner than Anthropic’s or OpenAI’s. I would not pin a Grok SWE-bench number in a production buying guide unless xAI publishes the exact result and harness details directly.

DeepSeek (Open Source)

  • DeepSeek’s public API docs currently anchor the line around DeepSeek-V3.2 and later V3.2-Exp updates.
  • The public API aliases deepseek-chat and deepseek-reasoner are the stable surfaces to watch.
  • I would not write as if DeepSeek V4 is current until DeepSeek actually ships and documents it in the public changelog.

DeepSeek remains the most important open-weight lane for operators who want self-hosting or a lower-cost secondary provider. But the factual way to frame it right now is: V3.2 is the public line, not a hypothetical V4.

What the Benchmarks Actually Say (April 2026)

Benchmarks are useful, but only if you respect the harness.

Coding (vendor-published numbers are not apples-to-apples):

  • Anthropic’s Claude Opus 4.6 launch page says its SWE-bench Verified result was averaged over 25 trials, and explicitly notes 81.42% with a prompt modification.
  • Anthropic’s Claude Sonnet 4.6 launch page says the reported Sonnet score is with thinking turned off, and explicitly notes 80.2% with a prompt modification averaged over 10 trials.
  • OpenAI’s GPT-5 launch materials reported 74.9% on SWE-bench Verified. I have not found an official GPT-5.4-specific SWE-bench figure I trust enough to pin here, so I do not present 74.9% as a GPT-5.4 fact.
  • xAI’s public docs are clear on Grok lineup and pricing, but I did not find a first-party xAI benchmark page I trust enough to cite for Grok SWE-bench here.
  • Google talks more about broad benchmark leadership, multimodal performance, and long-context work than about leading with a single SWE-bench narrative.

What I trust more than leaderboard screenshots:

  • Anthropic gives enough methodology notes to tell you when a score depends on prompt modifications, thinking settings, or Anthropic’s own scaffold.
  • OpenAI’s GPT-5 family clearly belongs in the default or coding lane, but exact model-version benchmark quoting should match the exact page OpenAI published.
  • Gemini keeps earning its place on long-context and multimodal workloads even when raw coding-benchmark discourse centers elsewhere.
  • xAI and DeepSeek can be excellent secondary lanes, but right now their pricing clarity is stronger than their benchmark clarity.

What Actually Happened With Anthropic and OpenClaw

This matters because it changed how every serious operator thinks about vendor dependency.

The facts, as reflected in Anthropic’s customer communication and OpenClaw’s Anthropic provider docs:

On April 4, 2026, Anthropic emailed Claude subscribers that effective immediately, they would “no longer be able to use your Claude subscription limits for third-party harnesses including OpenClaw.” Instead, this usage would require “Extra Usage” — pay-as-you-go billing separate from the subscription.

What Anthropic did NOT do:

  • They did not ban OpenClaw outright
  • They did not block the Anthropic API key path
  • Claude models are still fully available through the standard Anthropic API

What this means for operators:

OpenClaw’s own Anthropic provider docs are clear: the Anthropic API key remains the clearest, most predictable production path. No Extra Usage billing, no subscription ambiguity. If you run an API key, nothing changed for you.

What broke was the assumption that a consumer Claude subscription (Pro/Max) could power production-grade agent workflows through third-party tools indefinitely. That assumption was always fragile. On April 4, it became officially unsupported without Extra Usage.

The lesson that every operator should internalize:

If your whole stack depends on one vendor’s consumer subscription remaining compatible with your third-party tooling, you don’t have a production setup. You have a convenience that’s borrowing time.

Use the documented, explicit path. For Anthropic, that’s the API key. For OpenAI, that’s the API or the Codex subscription. For Google, that’s the Gemini API. Build on what the vendor tells you is the supported path, not on what happens to work today.

The Routing Table: What to Use for What

Here’s the concrete recommendation. Not “use the best model” — a routing table you can actually encode in OpenClaw.

OpenClaw already expects this kind of setup. The selection order is explicit in the docs: primary model first, then configured fallbacks, with auth-profile rotation happening inside the current provider before OpenClaw moves to the next model. It also gives you separate lanes for imageModel, pdfModel, and the generation tools. So stop thinking in terms of one favorite model. Think in terms of lanes.

LaneWhat belongs hereDefault pickEscalate / fallbackWhy this lane exists
Default operator laneSlack DMs, Telegram, normal research, tool use, coordinationClaude Sonnet 4.6 or GPT-5.4Escalate manually to Opus / GPT-5.4-pro for the hard stuffThis is your all-day lane. It needs to be smart, reliable with tools, and not expensive enough to punish normal usage.
Heavy reasoning laneArchitecture, policy decisions, difficult debugging, public-facing writing that must be rightClaude Opus 4.6 or GPT-5.4-proNo cheaper fallback inside the same job, this is already the premium laneUse this when the first wrong answer costs more than the token bill.
Coding laneParallel implementation agents, refactors, PR work, long-running code sessionsOpenAI Codex (openai-codex/gpt-5.4) or Claude Sonnet 4.6 via API keySame-provider sibling fallback, then a second providerCoding volume is where surprise bills and auth weirdness show up first. Give code its own budget and auth path.
Bulk / ops laneSummaries, extraction, nightly memory work, log triage, classificationGemini 2.5 Flash-Lite or GPT-5.4 nanoStep up to Flash / mini if quality is too weakCheap, fast, easy-to-verify work should not burn flagship-model money.
Media laneScreenshots, PDFs, images, video, multimodal understandingGemini 2.5 Pro or Gemini 3.1 ProSeparate imageModel / pdfModel / generation model if neededOpenClaw has dedicated config keys for this. Use them instead of forcing your text lane to do everything.
Budget secondary laneSecond opinions, overflow work, low-stakes experimentsGrok 4.1 Fast or DeepSeek V3.2Keep isolated from your core workflowUseful when cost matters more than polish, or when you want a cheap comparison pass.

A simple operator heuristic:

  • If the session is interactive and tool-heavy, optimize for reliability first.
  • If it runs on a cron, heartbeat, or batch pipeline, optimize for cost first.
  • If it touches screenshots, PDFs, images, audio, or video, route it to the media lane instead of pretending text-only defaults are enough.
  • If you’re about to run 5-20 copies in parallel, economics matter more than leaderboard bragging rights.

In OpenClaw config terms, make the table real:

  • Set agents.defaults.model.primary for the default lane.
  • Put deliberate fallbacks in agents.defaults.model.fallbacks.
  • Use agents.defaults.models for aliases and allowlisting, but remember the docs warning: once you set it, it becomes the model allowlist for /model and session overrides.
  • Set agents.defaults.imageModel or agents.defaults.pdfModel when your media lane differs from your text lane.

That’s what “route by job” looks like when it leaves the whiteboard and enters production.

The Fallback Rule: Why You Need More Than One Provider

Here’s the rule that separates operators who sleep well from operators who get paged at 2 AM:

Always have at least one cross-provider fallback configured before you need it.

Not “I’ll set one up if there’s a problem.” Not “I have the API key somewhere.” Configured, tested, ready.

The OpenClaw docs are very clear on the runtime order:

  1. Try the current model.
  2. Rotate auth profiles inside that provider if the error is failover-worthy.
  3. Only then move to the next model in agents.defaults.model.fallbacks.

That means same-provider and cross-provider fallbacks solve different problems.

Same-provider sibling fallback

Use this when you want to stay inside the same vendor and auth setup, but you don’t want one model outage or one model-scoped rate limit to take the whole lane down.

Example:

  • Primary: anthropic/claude-sonnet-4-6
  • Sibling fallback: anthropic/claude-opus-4-6

Why it helps:

  • OpenClaw may keep the same provider alive even when one model is cooling down.
  • Model-scoped rate limits can leave a sibling model usable.
  • You stay in the same provider family, so behavior and tool quirks are more predictable.

This is not enough on its own.

Cross-provider fallback

Use this when the problem is bigger than one model.

Example:

  • Primary: anthropic/claude-sonnet-4-6
  • Same-provider sibling: anthropic/claude-opus-4-6
  • Cross-provider fallback: openai/gpt-5.4
  • Cheap emergency lane: google/gemini-2.5-flash

Why it matters:

  • Provider-wide incidents do happen.
  • Billing and auth policy changes do happen.
  • A whole provider can be exhausted, disabled, or unusable even when your fallback model list looks good on paper.

The April 4 Anthropic change is the proof case. Operators who relied on one provider and one auth story had to scramble. Operators with a second provider already configured just kept moving.

The auth reality most people miss

OpenClaw doesn’t just fail over across models. It also manages auth profiles.

From the official docs:

  • Sessions pin the chosen auth profile to keep caches warm.
  • OpenClaw rotates profiles when needed, not on every request.
  • A user-pinned profile stays locked for that session, which is useful for debugging but removes some automatic within-provider recovery.

That matters because fallback is not only about models, it’s about auth shape too.

The practical provider guidance from OpenClaw’s official docs:

  • Anthropic: new setup should use an API key. The Claude subscription path inside OpenClaw is the one affected by Anthropic’s Extra Usage policy.
  • OpenAI: use the API or the Codex subscription/OAuth path. Both are explicit, supported routes in OpenClaw.
  • Google: use the Gemini API key path.

So a good fallback chain is not just “different model names.” It’s different vendors and, when possible, different auth stories.

A concrete config shape:

{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-sonnet-4-6",
        fallbacks: [
          "anthropic/claude-opus-4-6",
          "openai/gpt-5.4",
          "google/gemini-2.5-flash"
        ]
      }
    }
  }
}

Then verify it like an adult:

  • openclaw models status to confirm auth exists for every provider in the chain
  • /model status in chat when you want to confirm what the session is actually using

If your fallback path only exists in your head, you do not have a fallback path.

Five Rules That Survive Model Churn

Model names change. These principles don’t.

1. Encode the routing table in config, not in your memory. Use primary, fallbacks, aliases, and separate media lanes. If you use agents.defaults.models, remember it becomes the allowlist. A routing plan that isn’t actually selectable in OpenClaw is just a note to yourself.

2. Use the documented auth path, not the clever path. Anthropic API key, OpenAI API or Codex, Google API key. If the vendor treats a path as secondary, fragile, or extra-billed, don’t build your production setup on it.

3. Keep chat, coding, and media as separate lanes. They have different economics, different latency tolerances, and different failure modes. OpenClaw gives you separate knobs for a reason.

4. Put a different provider in the fallback chain, not just a bigger sibling. Same-provider siblings are useful, but they do not protect you from provider-wide auth, billing, or policy problems. Cross-provider fallback is the part that makes the system resilient.

5. Re-audit the stack on a schedule. This chapter has a date on it for a reason. Re-check pricing, model IDs, and auth health with the official docs and openclaw models list/status every 30-60 days. Model ops drift unless someone owns them.

Sources (Official Documentation Only)