AI Tools

Claude Opus 5.5 for Marketers: What Opus 5 Should Have Been on Day One (and the Routing Principle That Falls Out)

Claude Opus 5.5 for Marketers: What Opus 5 Should Have Been on Day One (and the Routing Principle That Falls Out)
Contents

I ran Opus 5.5 for the first 36 hours on the same workloads I'd put any new flagship through — a 9-touch B2B nurture, the daily Reddit monitoring agent, a 240-page competitor dump for the strategy brief due Friday, an SEO content refresh pipeline on a client's 80-post archive, and a 12-agent orchestration I had running overnight. By Monday morning I'd already started moving Opus 5 off three of those. That's not the reaction I expected to have two months after my Opus 5 take.

Anthropic shipped Claude Opus 5.5 on Tuesday, September 22, 2026. It's the first model in the new Claude 5.5 family. The headline framing: "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads." After a week with it, I think that framing undersells the upgrade in the wrong direction. The benchmark claim is mostly true. But the more important shift is that Opus 5.5 finally fixes what was wrong with Opus 5 — the verbose, hedged, "Claude-slop" writing; the mid-session drift on long outputs; the brutal cache-read bill on agentic loops.

In other words: Opus 5.5 is what Opus 5 should have been on day one. The mid-cycle update we deserved in August.

What Opus 5.5 is, briefly

The basics that matter for routing decisions:

  • API model id: claude-opus-5-5. Live in Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS.
  • Context: 1M input / 128K output. Same window as Opus 5 and Fable 5.1. No long-context surcharge.
  • Knowledge cutoff: June 2026. Same as Fable 5.1.
  • Pricing: $4 input / $20 output per million tokens. That's 20% under Opus 5's $5/$25 — and 60% under Fable 5.1's $10/$50. Cache reads dropped from Opus 5's $0.50 to $0.20 (60% cut), which is the number that actually changes agent economics.
  • Speed: Output is more than 30% faster than Opus 5. Fast mode (Claude API / Claude Code only) goes up to 2.5× at $8/$40.
  • Thinking: Adaptive, always on. Default effort is now medium, down from Opus 5's high. Thinking cannot be disabled — thinking: {"type": "disabled"} returns 400.
  • Latency tier: "Moderate." Faster than Fable 5.1 ("Slower") but slower than Sonnet 5 ("Fast").
  • Safeguards: First Opus model shipped with Fable-class cybersecurity and biology safeguards. Some cyber tasks now quietly re-route to Opus 4.8.

The pricing table that matters for the routing decision:

Model Input Output Cache read Default effort
Fable 5.1 $10 $50 $0.25 high
Opus 5.5 $4 $20 $0.20 medium
Opus 5 $5 $25 $0.50 high
Sonnet 5 $2 $10 $0.20 high

Read those rows carefully. Opus 5.5 is the new default-tier workhorse at the Opus capability level. Fable 5.1 stays on top for the hardest reasoning. Sonnet 5 and the eventual Sonnet 5.5 / Haiku 5.5 fill in below.

The benchmark card, with the footnotes that actually matter

Anthropic's launch table is the usual eight-row showcase. The honest read on it, after the noise settles:

  • Where Opus 5.5 beats Fable 5.1: Terminal-Bench 4.0 (66.4% vs 55.8%), FrontierCode v1.1 (54.4% vs 50.3%), CursorBench 4.0 (57.8% vs 51.8%), GDPval-AA v2.1 (1,846 vs 1,735 Elo), OSWorld 2.0 (81.8% vs 80.7%), Chartography (89% vs 88.4%). Artificial Analysis ranks Opus 5.5 #1 on its Intelligence Index at 58, five points ahead of Fable 5.1 and GPT-6 Astra.
  • Where Fable 5.1 still leads: A narrow band on Terminal-Bench-Science (64.6% vs 58.7%, mostly because safeguards kicked in and dragged Opus 5.5 down) and on the longest-context reasoning tasks Anthropic didn't put in the headline table.
  • Where GPT-6 Astra still leads: AutomationBench (41.4% vs 40%) and Terminal-Bench-Science. So the honest version of the claim is "best on coding and terminal work, not best at everything."

Anthropic says the gap to Fable 5.1 is "narrower than these scores suggest" in its own use. Independent reporting (eesel AI, MetricNexus, StationX) generally confirms the benchmarks at default medium effort, with one important caveat: at max effort, Opus 5.5 burns ~119K output tokens per task versus Astra's ~27K. That means on the same task, the bill on Opus 5.5 max can be $5.98 versus Astra's $3.26, even though the per-token price is lower. Reasoning effort is the dial you tune per step — leave max for the deliverables you actually need to be sharp.

The four jobs Opus 5.5 earns from Opus 5

These are the upgrades I'm already running in production.

1. The default model for agentic coding and research agents

Opus 5.5 at medium effort replaces Opus 5 at high effort for about a fifth of the cost per task on Anthropic's own Terminal-Bench 4.0 chart. The partner numbers are more concrete than the benchmark: GitHub reports finishing a multi-day rebase of 40 stacked pull requests in one Opus 5.5 session with all 40 passing CI. Quantium took a coding task that needed 38 prompts over four days on Opus 5 down to 11 prompts over three hours. Deloitte's code-review harness caught 72% of known bugs at Opus 5.5's lowest effort setting, against 56% for Opus 5 at high effort.

For marketers: this is the model to default your Lindy / n8n / Make / Claude Code orchestration to. The 60% cache-read cut is the part that matters — the 12-agent research workflow where Opus 5 was costing me ~$11 a session dropped to ~$4 on Opus 5.5 in my first run yesterday.

2. Long outputs that don't read like "Opus slop"

This is the upgrade most marketers will feel first. Opus 5's verbosity was the single most-quoted complaint in 2026: Reddit's 9to5grinder put it bluntly: "The verbosity and hallucinations from Opus 5 are entirely gone." Anthropic shipped a side-by-side example: when asked to diagnose a billing bug, Opus 5 buried the answer under three paragraphs of hedging; Opus 5.5 opens with "The extra drop is a bug in the billing refactor" and gives you the patch in the next sentence.

Box reports using a third of the tokens for the same job, with answers 40% less verbose. Walleye Capital says Opus 5.5 mostly solved their evaluation suite on the lowest setting — Opus 5 couldn't. Bun's Jarred Sumner summarized the week: "writing like Opus 4.6, coding like Fable 5.1."

For the long-tail jobs I run weekly — the campaign post-mortem template, the brand-voice analyzer output, the 9-touch B2B nurture, the NotebookLM competitor research brief — Opus 5.5 is the first Opus model in this lineage that I trust to land the first paragraph without an edit pass. That alone is worth the upgrade for a marketer who ships 30+ long-form pieces a month.

3. Agentic knowledge work at frontier cost

On the GDPval-AA v2.1 Elo benchmark (44 occupations, real work artifacts), Opus 5.5 sits at 1,846 vs Fable 5.1's 1,735 and GPT-6 Astra's 1,542. Anthropic's earnings-report test — give the model a copy of the web where the source is hard to find, then auto-grade every figure and quote — passed 16 of 18 Opus 5.5 reports. Neither Fable 5.1 nor Opus 5 passed in any attempt.

That's the model I want writing the strategy brief from a 240-page quarterly competitor dump, or turning the multi-agent competitive intel pipeline output into something the client can read without me rewriting the first paragraph.

4. The "60% cheaper" cache math for 200+ turn agents

This is the line item that quietly changes the agent math. On Opus 5, cache reads were $0.50 — 10% of the input price. On Opus 5.5, they're $0.20 — 5% of the input price. On any long-running agent that re-reads a brand-context document, a 400K-token archive, or a Screaming Frog crawl on every turn, the savings compound fast. A 200-turn session that cost me $4.10 in cache reads on Opus 5 costs $1.64 on Opus 5.5. Add the 20% input/output price cut and the per-task drop lands at the 40% Anthropic quotes.

The real win: jobs that were marginal before are now obviously worth running. The daily competitor-monitor session that I'd been running on Opus 5 every other day to save budget is back to every day.

Three jobs where Fable 5.1 or Opus 5 still wins

Don't let Opus 5.5 cannibalize the lanes where it doesn't actually win. After a week of testing, three jobs stay where they were:

1. The longest, broadest-context reasoning jobs — Fable 5.1

Fable 5.1 remains the right pick for jobs where you're reading a 1M-token pile in one session and asking the model to do the synthesis without chunking. Anthropic itself recommends Fable 5.1 "when Opus 5.5 at higher effort falls short." The use cases I keep on Fable 5.1: the 1,000-transcript customer-interview recall job, the full-quarter campaign synthesis with the post-mortem at the end, and any agent session that runs 500+ turns and needs the persistent context to stay sharp.

A reader on Reddit summarized the split well: "My expectation of where Fable will continue to shine is broad vague analysis across a broad range of domains. I expect Opus 5.5 will be an advancement in doing what Opus does well, not something else."

2. High-volume, low-stakes interactive — Sonnet 5 (and Opus 5 where you already have it)

For 50-variant RSA generation, 30 subject-line candidates, the Slack /research command, anything user-facing where the user is waiting on the response — Sonnet 5 at $2/$10 is the right default. Opus 5.5 isn't slow, but it isn't designed for this lane. And the eventual Sonnet 5.5 / Haiku 5.5 are where the cheap tier goes next; if you don't need the Opus-quality writing, you don't need to pay for it.

3. Regulated refusal and "I need it to push back" — Opus 5 or Fable 5.1

If your work is in regulated verticals where the model's calibration on refusal matters — financial advice, medical copy, anything EU AI Act gated — keep the older models until Opus 5.5 has 90 days of independent audit data. Anthropic shipped the Fable-class safeguards, but they're new on the Opus tier and the cyber / biology re-routing to Opus 4.8 isn't fully public-tested in customer deployments yet. Treat this as a watch item, not a blocker — but don't migrate regulated pipelines on day one.

The four breaking API changes to patch tonight

Anthropic published the migration guide on launch day. Read it. Then patch these.

  1. Thinking can no longer be disabled. thinking: {"type": "disabled"} and thinking: {"type": "enabled", ...} both return 400. Omit the thinking field and use the effort parameter to control depth. Default is now medium. If your pipeline hardcodes an old thinking config, every call will 400 in the morning.
  2. tool_choice: any and tool return 400. Same break as Fable 5.1 last month. Switch to auto + explicit instruction, or strict: true / structured outputs. Test all tool_choice paths tonight.
  3. Thinking blocks are model-bound and conversation-bound. Older models can't read 5.5 thinking blocks. If you pass reasoning history through older checkpoints, the request fails or reasoning drops. If your agent edits a previous turn, thinking for that turn is invalidated and re-passes need to start fresh.
  4. Computer use tool version changed. On Claude API and Google Cloud, Opus 5.5 needs computer_toolset_20260801; the older computer_20251124 returns 400. On Amazon Bedrock, computer_20251124 still works. If you run multi-cloud, this matters more than the others.

Plus a non-breaking behavior change worth knowing: text between tool calls now comes back inside thinking blocks with empty text at the default display setting. If your agent streams tool-call gaps to a user as "thinking..." progress updates, it goes quiet between tool calls until you set a display value that returns the text.

The routing principle, after a week

Opus 5.5 is the new default for everything in the Opus capability band. Fable 5.1 stays for the longest, broadest-context reasoning jobs. Opus 5 stays where you already have it pinned. Sonnet 5 carries the high-volume interactive tier.

In practice, this looks like:

  • Opus 5.5 at medium — agentic orchestration, the strategy brief from the 240-page competitor dump, the 9-touch B2B nurture, the Reddit monitor escalation layer, the SEO content refresh sweep.
  • Fable 5.1 at high — the 1,000-transcript synthesis, the multi-day agent session, the document production job from dense PDFs where you need the long-context recall to stay sharp.
  • Sonnet 5 (or Sonnet 5.5 when it ships) — RSA variants, subject lines, the Slack /research command, anything user-facing where latency matters more than peak quality.
  • Haiku 5 (eventually Haiku 5.5) — the cheap tier when either of the above is overkill.

The bit Anthropic got right that most "new model" coverage misses: Opus 5.5's 60% cache-read cut is bigger than the 20% input/output cut, and it changes which jobs are worth running on the Opus tier at all. The agent loops that were marginal on cost — daily monitors, weekly sweeps, every-Sunday 2 AM batch pipelines — are back on the table. That's the upgrade that actually moves the routing math.

And patch the four breaking API changes tonight. The 9 a.m. Slack alert will tell you whether you did.