What Anthropic actually shipped on July 24
On July 24, 2026, Anthropic released Claude Opus 5 — just under two months after Opus 4.8 landed on May 28. The pitch is unusually specific for a frontier launch: Anthropic says Opus 5 "comes close to the frontier intelligence of Claude Fable 5 at half the price," and it kept the price tag identical to the model it replaces.
That is the headline most coverage buried. Opus 5 costs $5 per million input tokens and $25 per million output tokens — the exact same rate as Opus 4.8 — while roughly doubling its agentic coding performance. There is no launch promo to expire, no price step-up scheduled. It ships as the default model on Claude Max and the strongest model available on Claude Pro, and it is live across Claude.ai, Claude Code, Claude Cowork, and the API under the model ID claude-opus-5.
We verified that pricing directly against the live OpenRouter feed rather than taking the announcement at face value, and it checks out: $5 in, $25 out, a 1,000,000-token context window, and a 128,000-token maximum output. For the full Claude lineup and how the tiers ladder up, see our Anthropic provider page.
That is the headline most coverage buried. Opus 5 costs $5 per million input tokens and $25 per million output tokens — the exact same rate as Opus 4.8 — while roughly doubling its agentic coding performance. There is no launch promo to expire, no price step-up scheduled. It ships as the default model on Claude Max and the strongest model available on Claude Pro, and it is live across Claude.ai, Claude Code, Claude Cowork, and the API under the model ID claude-opus-5.
We verified that pricing directly against the live OpenRouter feed rather than taking the announcement at face value, and it checks out: $5 in, $25 out, a 1,000,000-token context window, and a 128,000-token maximum output. For the full Claude lineup and how the tiers ladder up, see our Anthropic provider page.
The benchmarks, and which ones actually matter
Opus 5's benchmark sheet is lopsided in an interesting way — it wins the tests that measure novel reasoning and agentic follow-through by wide margins, and roughly ties on the saturated ones.
On Frontier-Bench v0.1, Anthropic's agentic coding evaluation, Opus 5 scores 43.3%. Fable 5 manages 33.7%, GPT-5.6 Sol 34.4%, and Opus 4.8 just 18.7%. That is not an incremental gain over its predecessor; it is more than double.
The ARC-AGI-3 result is the one that raised eyebrows. Opus 5 scored 30.2% — against 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8. The ARC Prize foundation independently verified 30.16% at high reasoning effort and confirmed Opus 5 as the highest-performing model on ARC-AGI-3 as of launch day, noting it solved five public demo environments no model had beaten before. On the older boards it reaches 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 at maximum effort.
On knowledge work, Artificial Analysis' GDPval-AA v2 puts Opus 5 at 1,861 Elo — ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736), and a 268-point jump over Opus 4.8's 1,593. SWE-bench Verified is effectively saturated at this tier: Opus 5 hits 96.0% to Sol's 95.0%. If you want the background on why these leaderboards disagree with each other so often, we wrote up which LLM leaderboards to trust and how quality scores are calculated.
On Frontier-Bench v0.1, Anthropic's agentic coding evaluation, Opus 5 scores 43.3%. Fable 5 manages 33.7%, GPT-5.6 Sol 34.4%, and Opus 4.8 just 18.7%. That is not an incremental gain over its predecessor; it is more than double.
The ARC-AGI-3 result is the one that raised eyebrows. Opus 5 scored 30.2% — against 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8. The ARC Prize foundation independently verified 30.16% at high reasoning effort and confirmed Opus 5 as the highest-performing model on ARC-AGI-3 as of launch day, noting it solved five public demo environments no model had beaten before. On the older boards it reaches 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 at maximum effort.
On knowledge work, Artificial Analysis' GDPval-AA v2 puts Opus 5 at 1,861 Elo — ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736), and a 268-point jump over Opus 4.8's 1,593. SWE-bench Verified is effectively saturated at this tier: Opus 5 hits 96.0% to Sol's 95.0%. If you want the background on why these leaderboards disagree with each other so often, we wrote up which LLM leaderboards to trust and how quality scores are calculated.
Opus 5 vs Fable 5: half the price, and it usually wins
Fable 5 is Anthropic's frontier model, released June 9, 2026 at $10 per million input and $50 per million output — precisely double Opus 5. The uncomfortable finding for Fable 5 is that Opus 5 beats it on most published numbers while costing half as much.
Artificial Analysis, which evaluated Opus 5 ahead of release, puts it narrowly at the top of its Intelligence Index: Opus 5 at max effort scores 61, effectively tied with Fable 5 at 60, ahead of GPT-5.6 Sol at 59 and Kimi K3 at 57 — at 26% lower cost per task than Fable 5. Opus 5 also leads on Frontier-Bench (43.3% vs 33.7%), GDPval-AA v2 knowledge work (1,861 vs 1,747 Elo), and AA-Briefcase agentic knowledge work, where its 1,720 Elo sits 146 points ahead of Fable 5. On CursorBench 3.2 it lands within half a point of Fable 5's peak, and it surpasses Fable 5 on OSWorld 2.0 computer use at roughly one-third the cost.
Fable 5 holds one genuine benchmark advantage: it edges Opus 5 on SWE-bench Pro, 80.0% to 79.2%, and Anthropic still recommends it for the longest multi-day autonomous agent runs where small per-step reliability gaps compound.
One point of confusion worth clearing up: Mythos 5 is not a separate, more powerful model. Per Anthropic's own platform documentation, Fable 5 and Mythos 5 are the same underlying model at the same $10/$50 price — the difference is that Mythos 5 ships without the safety classifiers, and is limited-release to vetted cybersecurity researchers and infrastructure providers through Project Glasswing. So when Anthropic says Opus 5 trails Mythos 5 on cybersecurity exploitation (and on SWE-bench Pro, 80.3%), that is a comparison against an unclassified build most teams cannot buy.
Put money on it. A workload of 5M input and 2M output tokens per day runs about $75/day on Opus 5 and $150/day on Fable 5 — over $27,000 a year in difference for a sub-one-point gap on a single benchmark. Run your own numbers on the API cost estimator, or see the full head-to-head at Opus 5 vs Fable 5.
Artificial Analysis, which evaluated Opus 5 ahead of release, puts it narrowly at the top of its Intelligence Index: Opus 5 at max effort scores 61, effectively tied with Fable 5 at 60, ahead of GPT-5.6 Sol at 59 and Kimi K3 at 57 — at 26% lower cost per task than Fable 5. Opus 5 also leads on Frontier-Bench (43.3% vs 33.7%), GDPval-AA v2 knowledge work (1,861 vs 1,747 Elo), and AA-Briefcase agentic knowledge work, where its 1,720 Elo sits 146 points ahead of Fable 5. On CursorBench 3.2 it lands within half a point of Fable 5's peak, and it surpasses Fable 5 on OSWorld 2.0 computer use at roughly one-third the cost.
Fable 5 holds one genuine benchmark advantage: it edges Opus 5 on SWE-bench Pro, 80.0% to 79.2%, and Anthropic still recommends it for the longest multi-day autonomous agent runs where small per-step reliability gaps compound.
One point of confusion worth clearing up: Mythos 5 is not a separate, more powerful model. Per Anthropic's own platform documentation, Fable 5 and Mythos 5 are the same underlying model at the same $10/$50 price — the difference is that Mythos 5 ships without the safety classifiers, and is limited-release to vetted cybersecurity researchers and infrastructure providers through Project Glasswing. So when Anthropic says Opus 5 trails Mythos 5 on cybersecurity exploitation (and on SWE-bench Pro, 80.3%), that is a comparison against an unclassified build most teams cannot buy.
Put money on it. A workload of 5M input and 2M output tokens per day runs about $75/day on Opus 5 and $150/day on Fable 5 — over $27,000 a year in difference for a sub-one-point gap on a single benchmark. Run your own numbers on the API cost estimator, or see the full head-to-head at Opus 5 vs Fable 5.
Opus 5 vs GPT-5.6 Sol: same input price, cheaper output
OpenAI released the GPT-5.6 family on July 9, 2026, and GPT-5.6 Sol is its flagship tier — alongside the mid-tier Terra ($2.50/$15) and budget Luna ($1/$6). Sol charges $5 per million input tokens, matching Opus 5 exactly, but $30 per million output against Opus 5's $25. Our OpenAI provider page tracks the whole ladder on live pricing.
That 17% output gap matters more than it looks, because output tokens are where the bill actually lives. Output is priced at 5x input on Opus 5 and 6x on Sol, so a chatty agent that generates long responses feels the difference immediately — a point we unpack in the output multiplier decides your LLM bill.
On capability, Opus 5 leads most of the shared board: SWE-bench Pro (79.2% vs 64.6%), ARC-AGI-3 (30.2% vs 7.8%), Frontier-Bench (43.3% vs 34.4%), OSWorld 2.0 (70.6% vs 62.6%), MCP Atlas (85.8% vs 75.3%), and GDPval-AA v2 (1,861 vs 1,736). Sol pushes back on Terminal-Bench 2.1 and BrowseComp, where it holds a narrow lead, and it ships a marginally larger 1.05M-token context window plus substantially faster peak throughput on specialised hardware.
The honest summary: Opus 5 is the better default, Sol is the better choice if your workload is terminal-heavy or latency-critical. Compare them line by line at Opus 5 vs GPT-5.6 Sol, and price the whole GPT-5.6 ladder in our OpenAI API pricing guide.
That 17% output gap matters more than it looks, because output tokens are where the bill actually lives. Output is priced at 5x input on Opus 5 and 6x on Sol, so a chatty agent that generates long responses feels the difference immediately — a point we unpack in the output multiplier decides your LLM bill.
On capability, Opus 5 leads most of the shared board: SWE-bench Pro (79.2% vs 64.6%), ARC-AGI-3 (30.2% vs 7.8%), Frontier-Bench (43.3% vs 34.4%), OSWorld 2.0 (70.6% vs 62.6%), MCP Atlas (85.8% vs 75.3%), and GDPval-AA v2 (1,861 vs 1,736). Sol pushes back on Terminal-Bench 2.1 and BrowseComp, where it holds a narrow lead, and it ships a marginally larger 1.05M-token context window plus substantially faster peak throughput on specialised hardware.
The honest summary: Opus 5 is the better default, Sol is the better choice if your workload is terminal-heavy or latency-critical. Compare them line by line at Opus 5 vs GPT-5.6 Sol, and price the whole GPT-5.6 ladder in our OpenAI API pricing guide.
Kimi K3: the open-weight model that undercuts everyone
Eight days before Opus 5, Moonshot AI shipped Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model that activates just 16 of 896 experts per token, roughly 1.8% of the network, with a 1,048,576-token context window, native vision, and always-on reasoning. Its Kimi Delta Attention is reported to decode about 6.3x faster than the previous generation.
It is far cheaper than anything Anthropic or OpenAI offers at the frontier: $3 per million input tokens on a cache miss, $15 per million output, and just $0.30 per million on cache hits. That last number is the real story for high-volume workloads with stable prefixes — see prompt caching can save 90% on AI costs for how to exploit it.
K3 took first place on LMArena's Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing. Moonshot is refreshingly candid about the rest: the company states K3 still sits behind Fable 5 and GPT-5.6 Sol on overall capability, even as it outperformed every other model in its own evaluation suite. Artificial Analysis backs that up — K3 scores 1,687 on GDPval-AA v2, third behind Opus 5 and Fable 5, and 57 on the Intelligence Index against Opus 5's 61.
The open-weights question is live as we publish. Moonshot committed to releasing full model weights on July 27, 2026, and its Hugging Face repository was still counting down at the time of writing. If those weights land as promised, K3 becomes a genuine self-hosting option in a way no Claude or GPT model is — check the repository directly before planning around it, and watch live K3 pricing on our Moonshot AI provider page.
The practical read: K3 is the value play, Opus 5 is the capability play. If your workload is high-volume and cost-sensitive, K3 at $3/$15 versus Opus 5 at $5/$25 is a 40% saving on both sides of the meter — before cache hits. If the hardest 10% of your requests determine whether the product works, that saving is false economy. We compare the trade-off directly at Opus 5 vs Kimi K3 and more broadly in open-weight vs proprietary LLM costs and our guide to newer providers like GLM and Kimi.
It is far cheaper than anything Anthropic or OpenAI offers at the frontier: $3 per million input tokens on a cache miss, $15 per million output, and just $0.30 per million on cache hits. That last number is the real story for high-volume workloads with stable prefixes — see prompt caching can save 90% on AI costs for how to exploit it.
K3 took first place on LMArena's Frontend Code evaluation at 1,679 points, ahead of Fable 5, in blind developer testing. Moonshot is refreshingly candid about the rest: the company states K3 still sits behind Fable 5 and GPT-5.6 Sol on overall capability, even as it outperformed every other model in its own evaluation suite. Artificial Analysis backs that up — K3 scores 1,687 on GDPval-AA v2, third behind Opus 5 and Fable 5, and 57 on the Intelligence Index against Opus 5's 61.
The open-weights question is live as we publish. Moonshot committed to releasing full model weights on July 27, 2026, and its Hugging Face repository was still counting down at the time of writing. If those weights land as promised, K3 becomes a genuine self-hosting option in a way no Claude or GPT model is — check the repository directly before planning around it, and watch live K3 pricing on our Moonshot AI provider page.
The practical read: K3 is the value play, Opus 5 is the capability play. If your workload is high-volume and cost-sensitive, K3 at $3/$15 versus Opus 5 at $5/$25 is a 40% saving on both sides of the meter — before cache hits. If the hardest 10% of your requests determine whether the product works, that saving is false economy. We compare the trade-off directly at Opus 5 vs Kimi K3 and more broadly in open-weight vs proprietary LLM costs and our guide to newer providers like GLM and Kimi.
The rest of the field: Gemini, DeepSeek and Grok
Opus 5 did not launch into a two-horse race. Three other providers matter for anyone pricing a frontier workload.
Google is the surprise absentee at the top. Despite months of speculation, there is no Gemini 3.5 Pro — Google shipped three Flash-class models on July 21 and explicitly no new Pro tier, leaving Gemini 3.1 Pro ($2 input / $12 output) as its current flagship reasoning model. The newest release, Gemini 3.6 Flash, runs $1.50/$7.50 with a 1M context. If you have seen 2M-context or Deep Think specs quoted for a "Gemini 3.5 Pro," those trace to rumour posts rather than Google.
DeepSeek remains the price floor. DeepSeek V4-Pro costs about $0.44 per million input and $0.87 per million output — roughly one-thirtieth of Opus 5 on output — while posting 80.6% on SWE-bench Verified, the strongest open-weights result and competitive with Gemini 3.1 Pro. For workloads where coding accuracy matters more than frontier reasoning, that ratio is hard to argue with.
xAI's Grok 4.5 lists at $2/$6 with a 500K context, but read the fine print: pricing doubles above 200K tokens, so the headline rate flatters it on long-context work. A flat comparison overstates how cheap it actually is at scale — the kind of trap we cover in why your LLM bill is higher than expected.
See every one of these against Opus 5 on live pricing with the price comparison tool.
Google is the surprise absentee at the top. Despite months of speculation, there is no Gemini 3.5 Pro — Google shipped three Flash-class models on July 21 and explicitly no new Pro tier, leaving Gemini 3.1 Pro ($2 input / $12 output) as its current flagship reasoning model. The newest release, Gemini 3.6 Flash, runs $1.50/$7.50 with a 1M context. If you have seen 2M-context or Deep Think specs quoted for a "Gemini 3.5 Pro," those trace to rumour posts rather than Google.
DeepSeek remains the price floor. DeepSeek V4-Pro costs about $0.44 per million input and $0.87 per million output — roughly one-thirtieth of Opus 5 on output — while posting 80.6% on SWE-bench Verified, the strongest open-weights result and competitive with Gemini 3.1 Pro. For workloads where coding accuracy matters more than frontier reasoning, that ratio is hard to argue with.
xAI's Grok 4.5 lists at $2/$6 with a 500K context, but read the fine print: pricing doubles above 200K tokens, so the headline rate flatters it on long-context work. A flat comparison overstates how cheap it actually is at scale — the kind of trap we cover in why your LLM bill is higher than expected.
See every one of these against Opus 5 on live pricing with the price comparison tool.
Effort settings, fast mode, and the costs nobody quotes
The sticker price is not the whole bill. Opus 5 ships with effort settings that let you trade intelligence against token consumption — the same control surface we analysed for the previous generation in effort control and the hidden token cost. Higher effort spends more reasoning tokens to reach better answers, and those tokens bill at the output rate. Anthropic's own ARC-AGI results show the spread: 88.3% at high effort versus 90.4% at maximum on ARC-AGI-2. Two points of accuracy, a materially larger bill.
There is also a separate Opus 5 Fast variant that runs about 2.5x faster at 2x the base price — $10 per million input and $50 per million output, which puts it at Fable 5 rates. It is a latency product, not a capability product.
Two new beta features change agent economics. Mid-conversation tool changes let you swap which tools Claude can access without invalidating the prompt cache, which previously forced an expensive full re-read of the context on every tool change. And automatic fallbacks route requests flagged by safety classifiers to another model instead of returning an error — relevant because Anthropic reports those classifiers intervene roughly 85% less often on Opus 5 than on Fable 5.
To see what any of this costs on your actual traffic, put your real numbers into the API cost estimator or convert a fixed budget into tokens with the token-to-USD calculator.
There is also a separate Opus 5 Fast variant that runs about 2.5x faster at 2x the base price — $10 per million input and $50 per million output, which puts it at Fable 5 rates. It is a latency product, not a capability product.
Two new beta features change agent economics. Mid-conversation tool changes let you swap which tools Claude can access without invalidating the prompt cache, which previously forced an expensive full re-read of the context on every tool change. And automatic fallbacks route requests flagged by safety classifiers to another model instead of returning an error — relevant because Anthropic reports those classifiers intervene roughly 85% less often on Opus 5 than on Fable 5.
To see what any of this costs on your actual traffic, put your real numbers into the API cost estimator or convert a fixed budget into tokens with the token-to-USD calculator.
Where Opus 5 is genuinely better, and where it is not
Beyond the headline coding numbers, Opus 5 posts real gains in scientific domains: Anthropic reports it beats Opus 4.8 on every life-sciences evaluation it ran, with organic chemistry up 10.2 percentage points and protein structure prediction up 7.7. It scored about 1.5x the pass rate of the next competitor on Zapier's AutomationBench, including a clean 100% on the account-health workflow.
It is also, by Anthropic's measure, the most aligned Opus model shipped — scoring 2.30 on its automated behavioural audit, its best result to date, with the lowest rate of deceptive behaviour and the strongest resistance to being manipulated into misuse.
The limitations are real and worth stating plainly. Opus 5 trails Mythos 5 substantially on cybersecurity exploit development and remains behind it on biology research. Anthropic keeps safeguards that permit source-code vulnerability identification while blocking binary scanning, penetration testing, and exploit generation. And on very long autonomous research runs, Anthropic flags meaningful limitations that argue for human checkpoints rather than unattended multi-day loops.
If your work sits in those zones, the frontier model is not automatically the right answer — why the cheapest LLM isn't always the best value cuts both ways.
It is also, by Anthropic's measure, the most aligned Opus model shipped — scoring 2.30 on its automated behavioural audit, its best result to date, with the lowest rate of deceptive behaviour and the strongest resistance to being manipulated into misuse.
The limitations are real and worth stating plainly. Opus 5 trails Mythos 5 substantially on cybersecurity exploit development and remains behind it on biology research. Anthropic keeps safeguards that permit source-code vulnerability identification while blocking binary scanning, penetration testing, and exploit generation. And on very long autonomous research runs, Anthropic flags meaningful limitations that argue for human checkpoints rather than unattended multi-day loops.
If your work sits in those zones, the frontier model is not automatically the right answer — why the cheapest LLM isn't always the best value cuts both ways.
Which model should you actually run?
For most teams, Opus 5 is now the default frontier model. It is strictly better than Opus 4.8 at identical pricing, so that upgrade is close to unconditional — same 1M context, same API shape, same rate, much better numbers. See Opus 5 vs Opus 4.8 if you need to justify it internally.
Against Fable 5, Opus 5 wins on price and on most benchmarks; keep Fable 5 only for multi-day autonomous agents where its SWE-bench Pro edge and long-horizon reliability earn the 2x. Against GPT-5.6 Sol, Opus 5 is cheaper on output and stronger on reasoning; Sol wins on terminal work and speed. Against Kimi K3, Opus 5 costs more but delivers more, and K3's open weights are the deciding factor if you need to self-host.
The best-value answer is usually not a single model at all. Route the bulk of traffic to a cheaper tier like Claude Sonnet 5 or Kimi K3 and escalate only the requests that clear a complexity threshold to Opus 5. Teams doing this typically cut spend 30–50% versus running a flagship everywhere, with no quality loss where it counts — the mechanics are in multi-model routing with quality scores and how to reduce AI API costs.
Before you commit to any of them, line all five up side by side on the price comparison tool, and price your real token distribution with the API cost estimator. The right model is rarely the most expensive one — it is the cheapest one that clears your quality bar.
Against Fable 5, Opus 5 wins on price and on most benchmarks; keep Fable 5 only for multi-day autonomous agents where its SWE-bench Pro edge and long-horizon reliability earn the 2x. Against GPT-5.6 Sol, Opus 5 is cheaper on output and stronger on reasoning; Sol wins on terminal work and speed. Against Kimi K3, Opus 5 costs more but delivers more, and K3's open weights are the deciding factor if you need to self-host.
The best-value answer is usually not a single model at all. Route the bulk of traffic to a cheaper tier like Claude Sonnet 5 or Kimi K3 and escalate only the requests that clear a complexity threshold to Opus 5. Teams doing this typically cut spend 30–50% versus running a flagship everywhere, with no quality loss where it counts — the mechanics are in multi-model routing with quality scores and how to reduce AI API costs.
Before you commit to any of them, line all five up side by side on the price comparison tool, and price your real token distribution with the API cost estimator. The right model is rarely the most expensive one — it is the cheapest one that clears your quality bar.