Reflection Beam vs Claude Opus 5.5: One You Can Call Today, One You Have to Wait For

Reflection Beam and Claude Opus 5.5 are not really rivals yet, and the reason is administrative rather than technical. Beam, Reflection's first model, was announced on 5 October 2026 and is a 501-billion-parameter sparse mixture-of-experts model activating 23 billion parameters per token — but it is reachable only through a waitlist, its weights are promised for later this month under Apache 2.0, and no price has been published anywhere. Opus 5.5 shipped on 22 September 2026 as Anthropic's flagship: a 1-million-token window, text, image and file input, and a list price of $4.00 per million input tokens and $20.00 per million output. So the honest version of this comparison is that one of these models you can put in production this afternoon with a known cost, and the other you cannot — and everything else here is a projection about what happens when the weights land.

The short answer

If the question is which model to build on this week, it is Opus 5.5 without much argument: it is generally available, independently scored, multimodal, priced, and served through an API you can call in five minutes. Beam's case does not rest on beating Opus 5.5 — nothing in Reflection's own material claims it does. It rests on a much narrower proposition: that at a given reasoning score, Beam burns roughly a quarter of the inference compute. If that holds up when somebody outside Reflection measures it, Beam becomes interesting for high-volume work where the answer has to be good but not best. If it does not, Beam is a mid-pack open model with a waitlist.

What follows separates the parts of this matchup that are facts today from the parts that are bets.

What each model is, precisely

Opus 5.5 is the closed, hosted successor to Claude Opus 5, positioned by Anthropic for multi-step changes across large codebases, code review, and long autonomous sessions. It accepts text, images and files, returns text, serves a 1,000,000-token context window with up to 128,000 output tokens, and exposes extended thinking with a configurable reasoning effort. It is not open weights, and its knowledge is bounded by Anthropic's training cutoff rather than by anything you control.

Reflection Beam is the opposite construction. It is 501 billion total parameters with 23 billion active — about 4.6 per cent of the model awake per token, an aggressive ratio even against GLM 5.2's roughly 5.4 per cent — built to be cheap to serve rather than cheap to train. It is text-only. Its API context is 256,000 tokens with a footnote warning the figure may move during the beta, and its maximum output is 128,000 tokens. The endpoint is OpenAI-compatible at api.reflection.ai/openai/v1 and exposes Chat Completions plus a Models list and nothing else. Its reasoning_effort parameter takes five values, defaults to medium, and cannot be turned off.

• Price — Claude Opus 5.5 $4.00 in and $20.00 out per million tokens, with cache reads at $0.20 and cache writes at $5.00, versus Reflection Beam: not published on the blog post, the documentation or the model listing.

• Availability — Opus 5.5 is generally available today; Beam is a waitlist for a beta, with the weights, technical report and model card promised later this month.

• Modality — Opus 5.5 takes text, images and files; Beam takes text only.

• Context — 1,000,000 tokens against 256,000, with Beam's number carrying a beta caveat.

• Open weights — no for Opus 5.5, promised under Apache 2.0 for Beam, not yet delivered.

• Independent score — Opus 5.5 carries one; Beam does not.

Price is the part that does not compare

Opus 5.5's $4.00 and $20.00 are the provider's list rate, and on our catalogue they are passed through unchanged with nothing added. Cache reads at $0.20 and cache writes at $5.00 matter enormously for the workload Opus 5.5 is sold into — a long agentic session that re-reads the same 200,000-token codebase turn after turn pays the cache rate on almost all of it, not the input rate. That is a real and often underestimated lever, and it is worth modelling before you conclude that a frontier model is out of budget.

Beam has no comparable number. There is no list price to compare against, no cache tier to model, and no published rate for the beta. Reflection has not said whether Beam will be sold per token, metered per seat, or given away with the weights. Anyone quoting a cost-per-million for Beam today is inventing it.

The practical consequence: the price comparison is not "Opus 5.5 is expensive and Beam is cheaper." It is "one of these has a price and the other has a date." For a team deciding today, that asymmetry settles more than the benchmarks do.

Benchmarks: the two numbers that overlap, and why they still do not compare

Reflection's launch tables do not include Opus 5.5, or any frontier closed model. Its comparison set is entirely open or semi-open: Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8-Max and DeepSeek V4.1 Flash. So any Beam-versus-Opus table has to be assembled by hand, and assembling it honestly means naming the harness differences rather than smoothing them.

There are exactly two benchmarks where both models have a published figure, and neither pairing is controlled.

• Humanity's Last Exam — Reflection reports Beam at 36.2 with no tools; Artificial Analysis records Opus 5.5 at 61.4. Two different harnesses, two different configurations — Opus 5.5's figure is not labelled no-tools — and a twenty-five-point gap that says less about the models than about the setup.

• SciCode — Reflection reports Beam at 49.7; Artificial Analysis records 66.9 for Opus 5.5. Same benchmark, same problem: one number is the vendor's own run, the other is a third-party harness.

Everything else does not overlap at all. Beam's table is built on Terminal-Bench v2.1, while the current Artificial Analysis index runs Terminal-Bench 4.0 — a different version of the same suite, which is why Beam's 80.1 and Opus 5.5's 59.6 on that board are not a comparison and should never be printed next to each other. On the index itself, Artificial Analysis puts Opus 5.5 at 57.6 in the Max / Default Fallback configuration, ranked fourth of the 147 models in that run. Beam has no index row: the model pages return 404 and the leaderboard carries no Beam entry as of this morning.

The one genuinely independent statement about Beam on the record is Artificial Analysis saying, on 5 October, that Reflection had given it access and that it was independently benchmarking the model — and that early indicators suggest Beam will be one of the most token-efficient open models it has seen for its level of intelligence. That is a strong sentence from the right source and it is still a sentence without a number. It is the number that will decide this matchup.

Where the efficiency claim actually bites

Reflection's argument is not that Beam is smarter than a frontier model. It is that Beam scores comparably to GLM 5.2 while using 3 to 4× less inference compute, and that the gap widens against models in the 2-trillion-parameter family where Qwen 3.8-Max sits. The chart behind it estimates generated FLOPs as twice the active parameter count times the mean generated tokens per attempt, and its own caption concedes that this excludes prompt prefill, attention and serving overhead.

Take that at face value and the interesting target is not Opus 5.5 at all — it is the middle of the market. Opus 5.5 competes on capability per task and wins on the tasks that need judgement: multi-step refactors, code review, work where a wrong answer costs more than the tokens. A model that is 4.6 per cent awake per token competes on cost per task and wins where volume dominates and the quality bar is "good enough and verified downstream." Those are different products, and a comparison that pits Beam against Opus 5.5 on a single scoreboard is measuring the wrong thing.

There is a second-order effect worth naming. If Beam's efficiency claim holds, its natural home is the kind of workload where you would otherwise spend a frontier model's tokens on a task it is overqualified for. That is a routing decision, not a model choice, and it is the reason the price question matters more than the benchmark question here: you cannot route on cost to a model with no cost.

Running them side by side, when Beam is callable

Opus 5.5 is on our catalogue today at $4.00 and $20.00 per million tokens, list price passed through with nothing added, and it sits alongside 200-plus other models behind a single OpenAI-compatible key at api.orcarouter.ai/v1. If you wanted to A/B a workload between a frontier model and a cheap open one, that is the shape of the setup: two model IDs, one key, no second contract and no code change — and automatic failover in the routing means a provider having a bad afternoon does not take your pipeline with it.

Beam cannot be part of that yet, and we are not going to pretend otherwise. Reflection Beam is not one of our routes, and we will not point you at whoever might resell it. When the weights land under Apache 2.0 you will be able to host it yourself and route to your own endpoint; until then the only path is Reflection's waitlist.

For a workload where a frontier model is genuinely required, the comparison to make is not against Beam. It is against everything else on the catalogue: GLM 5.3 at $1.26 and $3.96, Qwen 3.8-Max at $2.00 and $6.00, and DeepSeek V4.1 Flash at $0.15 and $0.60, all read live this morning. That is the tier Beam will land in if it prices competitively, and it is a much harder neighbourhood than the launch chart implies.

What would change this verdict

Three things, in order of how much they would matter.

An Artificial Analysis row for Beam. Not the announcement that one is coming — the row. If the index lands near Opus 5.5's 57.6 while the token-efficiency claim survives contact with a third-party harness, the middle of the market gets genuinely uncomfortable and Beam becomes the default for a lot of high-volume work. If it lands nearer the open-weights cluster in the low forties, Beam is a solid cheap model and nothing more.

A price. A model with no list rate cannot win a cost argument, because there is nothing to compare. If Beam comes in below the GLM 5.3 tier, the efficiency story becomes a budget story very quickly. If it comes in above DeepSeek V4.1 Flash's $0.15 and $0.60 without a capability edge, it will struggle for the volume work it was built for.

The weights. Until they exist, "open-weight" is a statement about a licence that has not been granted, and no one can check the FLOPs-per-score arithmetic against a model they can actually run. That is the check Reflection has invited and has not yet enabled.

Until at least one of those lands, the practical choice is not close. Opus 5.5 is the model you can reason about, cost out, and ship on. Beam is the model you watch.