Beam 1.0 vs Solar Mini 4: A Discount That Expires, Against a Price That Does Not Exist

Six days from now, Upstage's launch discount on Solar Mini 4 runs out. The promotion has run in two phases since the model went live on 22 September: a first round at $0.03 per million input tokens and $0.12 per million output that ended on 11 October, and the 50-per-cent-off round now in force at $0.05 input and $0.20 output, which ends at 00:00 UTC on 23 October 2026. After that the list rate of $0.10 and $0.40 applies. That is a dated commercial fact, and it is worth leading with, because the model on the other side of this comparison has the opposite problem: Beam 1.0, Reflection's 501.3-billion-parameter open-weight mixture-of-experts model released on 16 October 2026, has no published price at all. One of these two is a decision with a deadline. The other is a decision with an unknown in it. Neither is the obvious pick, and the reason comes down to what you are actually buying: a hosted endpoint with an expiry date on its price, or a 273 GiB download and four GPUs.

These are not rival products in any ordinary sense — one is a 35B MoE served at $0.10 per million tokens, the other is a 501B open-weight checkpoint you host yourself — but they land in the same budget conversation more often than the size gap suggests, because both are pitched at the cheap end of long-context work.

What the two of them are

• Solar Mini 4 — 35 billion total parameters with 3 billion active per token, a 512,000-token context window, up to 128,000 output tokens, and support for chat, reasoning, structured outputs and tool calling. English, Korean and Japanese, with Upstage describing it as fluent Korean, strong English and Japanese. Training cutoff February 2026. No open weights.

• Beam 1.0 — 501.3 billion total parameters with 23.4 billion active, 52 layers of which 3 are dense and 49 MoE, 256 experts, gated grouped-query attention mixed with sliding-window attention, and a 1M-token context on the model card with 131,072 tokens of maximum output on the API. Text only, English. Apache 2.0 weights in BF16, FP8 and NVFP4.

The size difference is real but the active-parameter difference is not as large as it looks: 3B against 23.4B is a factor of eight, not a factor of fourteen. What Solar Mini 4 gives up is total capacity and, more importantly, any ability to be hosted outside Upstage's infrastructure.

The two ways to pay, and why they do not compare

Solar Mini 4 is priced like a commodity and hosted like one. On our catalogue and on Upstage's own console the list rate is $0.10 per million input tokens, $0.01 per million cached input, and $0.40 per million output — figures that exclude 10 per cent VAT — and during the current promotion those become $0.05, $0.005 and $0.20. Those three numbers are the second phase of a phased promotion rather than the launch rate: the pricing page's own structured data shows the first phase running from 22 September to 11 October at $0.03, $0.003 and $0.12, which means anyone who adopted on day one is watching their rate rise twice, not once. It went live on the Upstage Console, in the Playground and as an on-premises deployment with no preview label and no waitlist, so there is nothing provisional about the availability. The promotion is the only provisional thing, and it is provisional by six days.

Beam 1.0 has a first-party API at api.reflection.ai with a model ID, an O​penAI-compatible surface, documented rate limits measured per organisation in requests and tokens per minute and per UTC day, and a 429 path that sets x-should-retry false once a daily allowance is spent. It does not have a per-million-token price. Reflection's announcement says the API is "free to start" and the documentation carries a full rate-limit page and no rate card, so the only priced way to run Beam today is to download it.

That download has a hardware floor, and it is the figure to put next to Solar Mini 4's rate. The recommended NVFP4 build is 273 GiB and the reference configuration is four N​VIDIA B300 SXM GPUs connected by NVLink — a single box, which is a far lower bar than a 501B model normally implies. FP8 is 468.72 GiB and runs on eight H200s. BF16 is 933.81 GiB and lands on eight B200s or B300s. The serving recipes keep the KV cache in FP8 E4M3 with a 64-token page and a 262,144-token default context — the full 1,048,576 is available, it is just not what the reference commands ship.

So the honest framing of "which is cheaper" is: 20 cents per million output tokens for as long as the promotion lasts and 40 cents after, against the amortised cost of four B300s plus power plus the engineering time to serve them. At low volume that is not close. At sustained high volume, on a workload that looks like Beam's, it eventually inverts — and the crossover point is a spreadsheet only you can fill in, because Reflection has not given you the hosted alternative that would make it easier.

The evaluation question neither vendor has settled

Solar Mini 4 has one independent measurement and Beam has one, and the comparison is more instructive than either number alone.

Artificial Analysis lists Solar Mini 4 at an Intelligence Index of 24.1, with a cost per completed task of $0.36 and output throughput of 204 tokens per second. Beam 1.0 sits at 30 on the same index version — v4.3.2 — ranked nineteenth of 120 models, with no price and no recorded speed because Artificial Analysis still lists "No publicly available API" for it. The verbosity figure is the interesting one: 79M output tokens across the index against a 130M median, tenth of forty-six. That is the independent version of Reflection's "roughly three times fewer inference-time reasoning tokens" claim, and the third-party evidence is weaker than the marketing but points the same way.

Read those two numbers together and the pattern is consistent with everything else here. The cheaper model is the slower and less capable one; the more capable model is faster per token but currently unpriceable. Upstage has published no benchmark table of its own for Solar Mini 4 — the changelog entry carries specs and no scores — and Reflection's benchmark table, which reports SWEBench Verified 81.6, Terminal Bench v2.1 80.1 and GPQA Diamond 90.6 on the BF16 checkpoint, is entirely vendor-reported and unreproduced.

The language question, which is where the decision usually lands

Solar Mini 4 speaks English, Korean and Japanese, and Upstage's phrasing — "fluent Korean with strong English and Japanese capabilities" — is the correct order for a Korean lab. Its training cutoff is February 2026, seven months stale at release, which is normal for the size and worth knowing before you ask it about anything recent. Beam 1.0 is English-first, its card says so, and it was not built with bilingual document work in mind.

For anyone processing Korean or Japanese text, that single line decides the comparison before cost enters it: Beam 1.0 is not a candidate, and the useful question becomes whether Solar Mini 4 at $0.10 per million is the right hosted tier or whether a larger routed model is worth the premium on the hard subset. Our catalogue carries the neighbouring tiers with published rates — Qwen3.8-Max at $2.00 and $6.00 per million tokens with a 1,000,000-token window, GLM-5.2 at $1.40 and $4.40, and Kimi K3 at $3.00 and $15.00 with 1,048,576 tokens. Routing the language work to one of those and the bulk tier to Solar Mini 4 is a configuration rather than a migration: one OpenAI-compatible key, the provider's list price passed through with nothing added, and automatic failover so a promotion expiring or a tier degrading is a routing change instead of an outage.

For an English-only workload the calculus is different, and Solar Mini 4's 512K window at a tenth of the flagship price is genuinely attractive for bulk summarisation and extraction. Beam 1.0's case is that you get a much larger model for the price of hardware you already amortise — and that the weights cannot be withdrawn, repriced, or rate-limited by anyone.

What to do before the 23rd

If Solar Mini 4 is in your architecture, the promotion's expiry is a real opportunity and a real trap in the same shape. Cost the workload at both numbers, because a 50 per cent discount that expires in six days is not a price, it is a promotion, and a budget built on $0.05 input that quietly becomes $0.10 is the kind of line item nobody notices until the quarter closes. If the workload survives at list price, the promotion is upside. If it only works at the discounted rate, you have six days to find the alternative.

If Beam 1.0 is in your architecture, the deadline is not the problem — the missing price is. The ten-day gap between announcement and weights closed; the gap between weights and a rate card is still open, and until it closes you are choosing between self-hosting and a beta endpoint with organisation-wide limits and no published cost. That is not an argument against trying it. It is an argument for trying it on a route you can leave.

The one thing both models share is that neither has been independently reproduced on the benchmarks their vendors care about. Solar Mini 4's single outside number is an index score; Beam 1.0's is the same index plus a token count. Everything else in both columns came from the party selling the model, and the promotion expiry is the only fact in this article that is guaranteed to still be true in a week.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.