Claude Opus 4.8 on Mac: Legacy API vs Opus 5.5

Source review: October 10, 2026. Claude Opus 4.8 remains available through the Claude API and OpenRouter, but Anthropic now marks it active (legacy). For a new U.S. coding integration, start by comparing Opus 5.5. Anthropic charges 20% less per standard input or output token for 5.5. These are cloud models, not downloads for Apple Silicon. Opus 4.8 specifications · Opus 5.5 specifications.

Model status, capabilities and API list rates

Item	Opus 4.8	Opus 5.5
Status	Active (legacy)	Current Opus
Direct Claude API ID	claude-opus-4-8	claude-opus-5-5
Context window	1M tokens	1M tokens
Standard maximum output	128K tokens	128K tokens
Input per million tokens	$5.00	$4.00
Output per million tokens	$25.00	$20.00
Cache read	$0.50	$0.20
Default effort	high	medium
Thinking when omitted	Off	Always on, adaptive
Forced tool_choice: any/tool	Accepted	HTTP 400; use auto or none
Minimum cacheable prompt	1,024 tokens	512 tokens
Enterprise Priority Tier	Supported	Not supported
5-minute cache write	$6.25	$5.00
1-hour cache write	$10.00	$8.00

These are U.S.-dollar API rates, not Claude subscription prices.

Important: The 1M-token context window is billed at the normal input rate even above 200K tokens; there is no hidden Opus long-context surcharge in the official Claude price list. Optional fast mode instead costs $10/$50 per million on Opus 4.8 and $8/$40 on Opus 5.5, and an explicitly requested U.S.-only inference region on the first-party Claude API adds a 1.1× token multiplier. Fast mode, U.S.-only inference, Batch and cache rates must not be silently mixed into the base-price comparison. A U.S. billing address alone does not select U.S.-only inference. Applicable sales tax or other taxes can be additional depending on billing arrangements. Anthropic’s price list provides Batch API discounts on standard input and output; those direct API discounts should not be assumed for every OpenRouter route.

Does legacy mean it is shutting down?

Not today. Anthropic lists Opus 4.8 as „Active (legacy)“ with retirement no sooner than May 28, 2027. That phrase is a minimum date, not a scheduled shutdown. Existing production apps can continue using 4.8 while running migration tests.

What actually happens on a Mac

Your Mac runs the editor, agent shell, tests and related tools. It does not hold Opus 4.8’s model weights or perform its inference. Requests go to an external inference provider. Mac RAM affects your local applications, not the size of the remotely hosted model.

This distinction matters for privacy as much as hardware planning. When a client includes proprietary code, documents or logs in an API prompt, that content leaves the Mac. Terminal access does not make a cloud agent offline. Restrict tool permissions and send only the files required for the task.

OpenRouter access: check the provider

OpenRouter’s ID is anthropic/claude-opus-4.8, not the direct Claude API’s claude-opus-4-8. The OpenRouter listing displays a 1M-token context window and a $5/$25 input/output rate per million tokens.

OpenRouter may route across several upstream providers. A U.S. account does not establish U.S.-only data processing; the selected provider, its processing terms and the contract determine those details. Check tool support, caching, context, data retention and billing before letting an agent send a private repository. Keep gateway keys in server-side secrets rather than a browser bundle or a Git commit.

How much does a single task cost?

Assume 100,000 ordinary input tokens and 10,000 output tokens, without caching, tools, Batch discounts or retries:

Model	Input	Output	Base token bill
Opus 4.8	$0.50	$0.25	$0.75
Opus 5.5	$0.40	$0.20	$0.60

This illustrates pricing, not relative quality or throughput. Long-running coding agents can issue many more requests and repeatedly send the same files. Cache-write and cache-read charges are different; it is incorrect to treat every prompt token as a discounted cache hit.

Migration from 4.8: real breaking changes in agent tools

The Anthropic migration guide identifies request errors and silent changes in agent output. Merely swapping the model name does not validate a migration:

Existing agent behaviour on 4.8	Required for Opus 5.5
Omits thinking or sets thinking: {"type":"disabled"}	Always-on adaptive thinking. Remove disabled/manual budgets; choose output_config.effort (low to max). Default drops from high to medium.
Enforces tool_choice: {"type":"any"} or a named tool	HTTP 400. Use tool_choice: {"type":"auto"} with schema validation or supported strict tools.
Displays content[0].text	Incorrect: the first block can be thinking. Select text blocks by type; use thinking.display: "summarized" for readable progress where appropriate.
Forgets the earlier assistant thinking block before sending tool_result	Send the complete assistant turn back unmodified or the API may return HTTP 400; do not silently drop signatures.
Uses computer_20251124 on the direct Claude API	Use computer_toolset_20260801. Partner-platform compatibility differs.
Relies on a 1,024-token minimum for prompt caching	Opus 5.5 can cache at 512 tokens, potentially changing cache eligibility and spend.
Depends on enterprise Priority Tier	Opus 5.5 does not support Priority Tier; plan capacity separately.

A second, non-fatal difference matters for agent products: narration between tool calls moves into thinking blocks. Default thinking display is omitted; a progress indicator that simply looks for interleaved text can appear to freeze. Preserve the original signed blocks during tool loops and adapt what the UI renders. This is especially relevant to multi-agent orchestration and reviewer loops.

Real cost of caching: account for each token category

For an illustrative 200K five-minute cache write, 200K cache read, 100K uncached input and 10K output (across the relevant requests), the direct Anthropic API list rates produce:

Billed component	Opus 4.8	Opus 5.5
Uncached input	$0.50	$0.40
5-minute cache write	$1.25	$1.00
Cache read	$0.10	$0.04
Output	$0.25	$0.20
Illustrative total	$2.10	$1.64

Those totals require real cache creation and hits; not every repeated prompt qualifies. One-hour cache writes are billed at a different rate. Hidden adaptive-thinking tokens also count as output; the same task may produce different output-token counts. No direct paid model call was run for this article.

Small billable API compatibility check (Python standard library)

This script makes one real, billable request per model when executed. It explicitly enables adaptive thinking and low effort on both to avoid a misleading default-effort comparison; the agent should still re-test other effort levels later. It prints the HTTP error status without exposing the key, counts response blocks by type, and prices uncached input, output, cache hits and both cache-write TTLs from the API’s usage fields. The calculation is a list-price estimate, excluding fast mode, region modifiers, Batch, server tools, discounts and tax.

Set ANTHROPIC_API_KEY as a secret in your environment and never commit it.

import json
import os
from urllib.error import HTTPError
from urllib.request import Request, urlopen

key = os.environ.get("ANTHROPIC_API_KEY")
if not key:
    raise SystemExit("Set ANTHROPIC_API_KEY in your environment")
rates = {
    "claude-opus-4-8": (5, 25, 0.50, 6.25, 10),
    "claude-opus-5-5": (4, 20, 0.20, 5, 8),
}
task = "Write a Python function to merge two sorted integer lists; include three asserts."

for model, (r_in, r_out, r_read, r_write5, r_write1) in rates.items():
    payload = {
        "model": model, "max_tokens": 4096,
        "thinking": {"type": "adaptive"},
        "output_config": {"effort": "low"},
        "messages": [{"role": "user", "content": task}],
    }
    request = Request(
        "https://api.anthropic.com/v1/messages",
        data=json.dumps(payload).encode("utf-8"),
        headers={
            "x-api-key": key,
            "anthropic-version": "2023-06-01",
            "content-type": "application/json",
        },
    )
    try:
        with urlopen(request, timeout=120) as response:
            result = json.load(response)
    except HTTPError as error:
        raise SystemExit(f"{model}: HTTP {error.code} - check model access and migration settings") from error

    usage = result["usage"]
    writes = usage.get("cache_creation") or {}
    w5 = writes.get("ephemeral_5m_input_tokens", 0)
    w1 = writes.get("ephemeral_1h_input_tokens", 0)
    if usage.get("cache_creation_input_tokens", 0) != w5 + w1:
        raise SystemExit("Unrecognized cache-write breakdown: check the actual invoice")
    plain = usage["input_tokens"]
    read = usage.get("cache_read_input_tokens", 0)
    output = usage["output_tokens"]
    estimate = (
        plain * r_in + read * r_read + w5 * r_write5
        + w1 * r_write1 + output * r_out
    ) / 1_000_000
    answer = "".join(block.get("text", "") for block in result["content"] if block["type"] == "text")
    print(model, "stop_reason:", result.get("stop_reason"))
    print("uncached/read/write5/write1/output:", plain, read, w5, w1, output)
    print("estimated base USD:", round(estimate, 6), "text chars:", len(answer))

A/B acceptance test: Run five real, non-sensitive bug fixes and five multi-file refactors with identical permissions. Measure test passes, retries, interruption/refusal handling, missing tool output, tokens, latency and billed cost per successful task, not only average tokens. Keep the original API/model pair available until the new agent passes that suite.

What should U.S. developers choose?

For an existing validated workflow, it is reasonable to retain 4.8 while testing a migration. For a new project, Opus 5.5’s lower API prices and current support make it the default comparison. If code must remain offline, choose a local open-weight model; no hosted Opus API is equivalent to local inference.

Primary sources reviewed October 10, 2026: Anthropic Opus 4.8, Opus 5.5, API pricing, OpenRouter.