BuildNot yet confirmed elsewhere1 publisher2 min readPublished
Claude 4.5 Opus builds sub-agent spawning into the model, according to a dev.to write-up
Anthropic launched Claude 4.5 Opus on October 9 with sub-agent spawning in the model and a 500,000-token context, per a dev.to write-up. The claimed SWE-bench Verified lead over GPT-6.1 Sol and Gemini 2.5 Ultra comes from that post alone.
The Engineer · Build desk
Five key figures. SWE-bench Verified score: 74.8 percent, with no rival scores printed in the post. Native context window: 500,000 tokens. Input list price: $15 per million tokens. Cost to fill the window once: $7.50 at the input rate, a desk calculation. Prompt caching discount: up to 85 percent.
- 74.8 percent SWE-bench Verified score, claim 2 Post prints no rival scores
- 500,000 tokens Native context window, claim 5
- 15 $/M tokens Input list price, claim 12
- 7.5 USD Cost to fill the window once, claim 17 Desk calculation at the input rate
- 85 percent Max prompt caching discount, claim 13 Repeated prompts and documents
| Figure | Value | Note | Claim |
|---|---|---|---|
| SWE-bench Verified score | 74.8 percent | Post prints no rival scores | 2 |
| Native context window | 500,000 tokens | 5 | |
| Input list price | 15 $/M tokens | 12 | |
| Cost to fill the window once | 7.5 USD | Desk calculation at the input rate | 17 |
| Max prompt caching discount | 85 percent | Repeated prompts and documents | 13 |
What happened
- A coordinator agent splits a directive into tasks, hands each to a specialised instance in an isolated sandbox, and ends dead-end paths before the token budget runs out.
- When standard APIs fall short, the model writes and compiles its own Python or Rust helper, runs it in an isolated micro-VM, and checks the parsed output against a schema.
- The post gives 74.8 percent as the SWE-bench Verified score on real-world GitHub repositories and prints no scores for GPT-6.1 Sol or Gemini 2.5 Ultra.
- List price is $15 per million input tokens and $75 per million output tokens.
Why it matters
- decision Teams that run an external harness face a pilot decision: run native sub-agents against their own issue backlog, with the harness as the control, before anything is retired.
- cost Filling the 500,000-token window once costs $7.50 at the input rate. If the 85 percent cache discount applies to input, a cached million tokens costs $2.25. Fan-out multiplies either figure by the number of sub-agents that load their own context.
- contradiction The post's lead paragraph promises "perfect factual recall" at 500,000 tokens, while its spec line says needle-in-a-haystack accuracy reaches "up to 100 percent".
According to the post, the spawning, monitoring and debugging primitives sit inside the core reasoning loop. The post contrasts that with earlier generations, which needed "fragile external prompt chaining frameworks or complex third-party agent harnesses" [8]. Each sub-agent gets an isolated sandbox [9], and synthesized tools run in an isolated micro-VM [10]. A model can decide to spawn. Something outside the weights has to start the sandbox, meter it and tear it down. The post does not say who hosts those sandboxes. If Anthropic does, the harness shrinks. If the customer does, we think it mostly moves: the code that starts the sandbox, holds the git credentials and logs the run still gets written, around a different call.
Tool synthesis changes what an allowlist looks like. The post says the model writes Python or Rust helpers when standard APIs fall short [10], and that this removes the need to pre-define hundreds of static tool declarations in system prompts [15]. A static declaration list is also an inventory of what the model can touch. Generated code replaces that inventory with a sandbox policy. The post says Anthropic updated its Constitutional AI filters to prevent autonomous shell injection and unverified network calls [16].
"Autonomous software engineering requires models that do not merely generate code snippets, but orchestrate full development lifecycles with continuous verification and architectural rigor," Anthropic's Head of Frontier Systems said, according to the post [14]. That sentence states a design goal.
A SWE-bench Verified score is a claim about one set of GitHub issues run under one scaffold [2]. For it to transfer, your issues have to resemble that set, your token budget has to match the eval's, and the scaffold has to be one you would actually run. If the eval used an external scaffold, the score does not measure replacing one. If it used native sub-agents, it is the first data point for that claim. Which case applies is the first thing to establish, and the comparison with GPT-6.1 Sol and Gemini 2.5 Ultra [3] cannot be checked until it is.
The only account here is a secondary one. The post's text closes with an unrelated block of advice on quantization and GPU junction temperatures, and it stops mid-word, at "reprodu" [1].
What to watch
- Anthropic's own documentation or model card confirming the 74.8 percent score, the scaffold used, and the GPT-6.1 Sol and Gemini 2.5 Ultra comparison scores.
- An independent run that reports tokens spent per resolved issue with native sub-agents on, against the same tasks under an external harness.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence15
- Adoption
- Insufficient
- Hype gap+60
- Incentives
- Insufficient
- Confidence25
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The post's text ends with an unrelated block of advice on quantization and GPU junction temperatures and stops mid-word, at "reprodu".
- [2]
The post gives 74.8 percent as the SWE-bench Verified score for autonomous issue resolution across real-world GitHub repositories, and prints no scores for GPT-6.1 Sol or Gemini 2.5 Ultra.
- [3]
The post says Claude 4.5 Opus surpasses OpenAI's GPT-6.1 Sol and Google's Gemini 2.5 Ultra on standard SWE-bench Verified evaluations.
- [4]
The dev.to post says Anthropic launched Claude 4.5 Opus on October 9, 2026.
ReportedInsufficientSource: dev.to post2 sources— create a free account to open themView cited source - [5]
The post says Claude 4.5 Opus has a 500,000-token native context window.
ReportedInsufficientSource: dev.to post2 sources— create a free account to open themView cited source - [6]
The post's lead paragraph describes the 500,000-token window as having "perfect factual recall", while its spec line lists bidirectional needle-in-a-haystack verification "up to 100 percent accuracy".
ReportedInsufficientSource: dev.to post2 sources— create a free account to open themView cited source - [7]
The post says Claude 4.5 Opus is available via the Anthropic API, Amazon Bedrock and Google Cloud Vertex AI.
ReportedInsufficientSource: dev.to post2 sources— create a free account to open themView cited source - [8]
The post says Claude 4.5 Opus includes architectural primitives for spawning, monitoring and debugging subordinate sub-agents directly within the core reasoning loop, unlike earlier generations that required "fragile external prompt chaining frameworks or complex third-party agent harnesses".
- [9]
The post says a primary coordinator agent decomposes directives into tasks, assigns isolated sandboxes to specialized instances, and terminates dead-end execution paths before token budgets are exhausted.
- [10]
The post says the model can write, compile and execute custom Python or Rust helper tools when standard APIs fall short, running them in an isolated micro-VM and validating parsed data against expected schemas.
- [11]
The post lists reasoning budget controls: API parameters that let developers cap inference compute by task complexity.
- [12]
The post says Anthropic priced Claude 4.5 Opus at $15 per million input tokens and $75 per million output tokens.
- [13]
The post says prompt caching discounts offer up to an 85 percent price reduction for repeated system prompts and document repositories.
- [14]
"Autonomous software engineering requires models that do not merely generate code snippets, but orchestrate full development lifecycles with continuous verification and architectural rigor," Anthropic's Head of Frontier Systems said, as quoted in the post.
ReportedInsufficientSource: Anthropic's Head of Frontier Systems, quoted in the dev.to postView cited source - [15]
The post says dynamic tool synthesis eliminates the need for developers to pre-define hundreds of static tool declarations in their system prompts.
- [16]
The post says Anthropic updated its Constitutional AI filters to prevent autonomous shell injection and unverified network calls.
- [17]
Filling the 500,000-token window once costs $7.50 at the $15 per million input rate.
- [18]
If the 85 percent maximum cache discount applies to the input rate, a million cached input tokens would cost $2.25.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toAnthropic Unveils Claude 4.5 Opus: Multi-Agent Orchestration and Autonomous Tool Synthesis
1 article · October 12, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Frontier model releasesFollow
- AI Coding AgentsFollow
- LLM benchmark resultsFollow