Claude Opus 4.5 Retains Reasoning Steps, Avoiding Forgetfulness in Long Tasks
When you’re pulling together a dozen code snippets or sifting through a dense research paper, a model that can actually remember its own reasoning feels almost like a safety net. The newest benchmark throws Claude Opus 4.5 against GPT-5.1 and Gemini 3 Pro, zeroing in on long, multi-step tasks. The headline claims “Claude Opus 4.5 Retains Reasoning Steps, Avoiding Forgetfulness in Long Tasks,” but the real question is whether it can still recall a decision made five turns ago without looping back into a dead end.
Earlier versions of these models tended to lose the thread once a conversation stretched past a few prompts, and that loss often made them repeat ideas that had already proven useless. In this head-to-head we’re looking to see if Opus 4.5 can break that habit, keeping its internal “thinking blocks” intact as the dialogue moves forward. From what the quote below suggests, there’s a clear shift in how the model treats its past reasoning - and that could be a game-changer for anyone stitching together complex workflows.
In long chains of coding or research, older models often forgot why they made a choice a few turns earlier. Opus 4.5 keeps its own thinking blocks intact from one step to the next. This stops it from repeating the same failed ideas.
It behaves like someone who remembers previous attempts with clarity. The model can zoom into small portions of a screen at full resolution. It uses this to catch minute details in documents or interfaces that other models miss.
A developer compared GPT-5.1, Gemini 3.0 and Opus 4.5 across three coding tasks to see how they behave in real work.
The three releases this month have kind of reset what we expect from big models when they’re asked to work on something that stretches over many turns. First came GPT-5.1, then Gemini 3 Pro, and finally Claude Opus 4.5 wrapped things up. The big difference is that they don’t stick to a single, straight-line train of thought like the older versions did.
Those earlier models would often forget why they made a choice a few steps back, and you’d see the same slip-ups pop up again. Opus 4.5, on the other hand, seems to keep its own reasoning chunks from one exchange to the next, so the same dead-end ideas don’t keep resurfacing. It feels a bit like a researcher who can actually remember what they’ve already tried - something that could make long-form coding or deep-dive investigations smoother.
I’m not sure yet whether this memory-style approach will consistently beat the old ways across all fields. The move to non-linear processing is definitely a shift, but we’ll have to watch how it plays out in real-world tasks. For now, the early signs point to a small boost in continuity, though broader testing is still on the horizon.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI OpenAI says AI saves knowledge workers 40‑80 minutes; use yields five‑fold gains
LLMs & Generative AI Grok Chat: AI for debugging, building, testing web apps with voice and images
LLMs & Generative AI Samsung adds Vision AI Companion, an AI Bixby, to TVs for real‑time queries
Related Topic Anthropic finds strict anti-hacking prompts increase AI sabotage and lying
Research & Benchmarks
Related Topic Tool Extracts Detailed Claude Code Transcripts via Reverse‑Engineered API
Open Source
Common Questions Answered
How does Claude Opus 4.5 prevent forgetfulness in long, multi‑step coding tasks compared to older models?
Opus 4.5 retains its own reasoning blocks from one turn to the next, so it remembers why a particular coding choice was made. This prevents the model from repeating failed ideas and keeps the development flow coherent across dozens of code snippets.
What advantage does Claude Opus 4.5 show over GPT‑5.1 and Gemini 3 Pro when analyzing extended research papers?
In the benchmark, Opus 4.5 not only keeps track of earlier reasoning steps but also can zoom into small screen portions at full resolution, catching minute details that other models miss. This combination lets it maintain context over many paragraphs while extracting fine‑grained information.
What does the article mean by “Opus 4.5 keeps its own thinking blocks intact from one step to the next”?
The phrase describes the model’s ability to store each intermediate reasoning segment as a discrete block, preserving the logic behind decisions made several turns earlier. By doing so, Opus 4.5 avoids losing the rationale behind actions, which older models often did.
How have GPT‑5.1, Gemini 3 Pro, and Claude Opus 4.5 reshaped expectations for large models handling extended work?
All three releases move away from a single, linear chain of thought, allowing more flexible, multi‑step processing. Claude Opus 4.5 further raises the bar by explicitly retaining reasoning steps, reducing repeat errors and improving performance on long‑duration tasks such as complex coding or deep research.