Claude is priced in a clean ladder — Fable 5.1 at the top, then Opus 5, Sonnet 5 and Haiku 4.5 — and the spread from top to bottom is 50x. This page puts Anthropic list prices next to what APIMODELS charges on each rung, including the cache columns most comparison pages leave out, and says plainly where going direct is still the better call.
A correction worth making up front, because a lot of pricing pages are still wrong about it: Claude Sonnet 5 lists at $2 / $10 per 1M input/output tokens. The $3 / $15 figure that circulated widely was a scheduled increase for 2026-09-01 that Anthropic explicitly cancelled — its own pricing page now states that $2 / $10 "is now the standard price" and the increase "will not occur." Any comparison built on $3 / $15 overstates every discount it claims against Sonnet 5, ours included until we caught it.
On APIMODELS, Fable 5.1 is $5 / $25 — exactly half Anthropic’s $10 / $50. Opus 5 is $3 / $15 against a $5 / $25 list, 40% below. Sonnet 5 is $1.60 / $8.00 against $2 / $10, 20% below. Haiku 4.5 is $0.353 / $1.765 against $1 / $5, about 65% below. The discount is not uniform across the ladder and we do not pretend otherwise — it reflects what each tier actually costs us upstream.
Prompt caching is where the ladder gets interesting, and where a headline discount can mislead. Anthropic reads cache hits at 0.1x base input on most models but at 0.025x on Fable 5.1, which puts its official cache read at just $0.25 per 1M. Ours is $0.22 — cheaper, but only by about 12%, not by the 50% that applies to Fable 5.1’s base rates. On Sonnet 5 the picture reverses: official cache read is $0.20 and ours is $0.10, a full 50% below. If your workload is cache-heavy, compare that column specifically rather than the headline.
Reliability, measured rather than claimed: over the last 30 days of external production traffic (internal accounts excluded), Sonnet 5 served 1,509 calls at a 97.7% success rate with a median of 8.2 seconds, and Opus 5 served 500 calls at 97.4% with a median of 8.8 seconds. Opus 5’s p95 is 283 seconds, which is what long agentic turns actually look like — set client timeouts accordingly. Failed requests are never billed. Fable 5.1 launched on 2026-09-03 and does not yet have a sample worth quoting, so we are not quoting one.
These run on the native Anthropic Messages API at /v1/messages, not an OpenAI-shaped translation layer. That means Claude Code, the Anthropic SDK and Cursor work after changing only the base URL and key — tool use, thinking blocks and streaming events keep their native shapes. No Anthropic account is needed, it is reachable from mainland China, and every model on the platform draws on one USD balance.
Price per 1M tokens — Anthropic list vs APIMODELS, verified 2026-09-05
input output cache read vs list
Claude Fable 5.1
Anthropic (list) $10.00 $50.00 $0.25
APIMODELS $5.00 $25.00 $0.22 -50%
Claude Opus 5
Anthropic (list) $5.00 $25.00 $0.50
APIMODELS $3.00 $15.00 $0.391 -40%
Claude Sonnet 5
Anthropic (list) $2.00 $10.00 $0.20
APIMODELS $1.60 $8.00 $0.10 -20%
Claude Haiku 4.5
Anthropic (list) $1.00 $5.00 $0.10
APIMODELS $0.353 $1.765 -- -65%
Note on the cache column: Anthropic reads cache hits at 0.1x base input,
except on Fable 5.1 where it is 0.025x. That is why Fable 5.1's official
cache read ($0.25) is already low and our edge there is ~12%, not 50%.
Sonnet 5 lists at $2/$10. The $3/$15 figure still quoted in many places
was a 2026-09-01 increase that Anthropic cancelled.cURL
# Native Anthropic Messages API — only the base URL and key change
curl https://api.apimodels.app/v1/messages \
-H "Authorization: Bearer $APIMODELS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "claude-sonnet-5",
"max_tokens": 512,
"messages": [{"role": "user", "content": "Refactor this function for readability."}]
}'Python
import anthropic
client = anthropic.Anthropic(
api_key="YOUR_APIMODELS_KEY",
base_url="https://api.apimodels.app",
)
msg = client.messages.create(
model="claude-fable-5-1", # or claude-opus-5 / claude-sonnet-5
max_tokens=1024,
messages=[{"role": "user", "content": "Plan a migration from Postgres 14 to 17."}],
)
print(msg.content[0].text)
print(msg.usage) # cache_read_input_tokens bills at the cache rate aboveAnthropic list prices per 1M input/output tokens are $10 / $50 for Fable 5.1, $5 / $25 for Opus 5, $2 / $10 for Sonnet 5 and $1 / $5 for Haiku 4.5. On APIMODELS the same models are $5 / $25, $3 / $15, $1.60 / $8.00 and $0.353 / $1.765 — about 50%, 40%, 20% and 65% below list respectively.
$2 / $10. The $2 / $10 rate launched as introductory pricing through 2026-08-31, and a rise to $3 / $15 was scheduled for 2026-09-01 — but Anthropic cancelled it and its pricing page now says $2 / $10 "is now the standard price" and the increase "will not occur." A lot of comparison content still quotes $3 / $15, which inflates whatever discount that page claims.
Yes. These models are served on the native Anthropic Messages API at /v1/messages, not an OpenAI-shaped translation, so Claude Code, the Anthropic SDK and Cursor work after pointing the base URL at https://api.apimodels.app and swapping the key. Tool use, thinking blocks and streaming event shapes are unchanged.
Cache reads bill at $0.22 per 1M on Fable 5.1, $0.391 on Opus 5 and $0.10 on Sonnet 5. Worth knowing: Anthropic reads cache hits at 0.1x base input on most models but 0.025x on Fable 5.1, so its official cache read is already only $0.25 and our advantage there is about 12% rather than the 50% that applies to its base rates. On Sonnet 5 the official cache read is $0.20 and ours is $0.10, a full 50% below. Cache token counts appear in every response’s usage object.
Three honest cases. First, the Batch API: Anthropic discounts asynchronous batch jobs by 50%, which we do not resell, so large offline workloads can be cheaper direct. Second, account-level products we do not carry — Fast mode on Opus 5, priority tiers, data-residency (inference_geo) commitments. Third, enterprise compliance paperwork such as DPAs and SOC 2 chains, which only a direct Anthropic, Bedrock or Vertex relationship provides.
Over the last 30 days of external traffic with internal accounts excluded, Sonnet 5 served 1,509 calls at 97.7% success (median 8.2s, p95 85.6s) and Opus 5 served 500 calls at 97.4% (median 8.8s, p95 283.6s). That p95 on Opus 5 is long agentic turns, not stalls — set client timeouts above 300 seconds for agent workloads. Failed requests are never billed. Fable 5.1 launched 2026-09-03 and we will publish its numbers once the sample is large enough to mean anything.
