Flash-1017 vs Opus 5 vs Sol on Price and Scores

DeepSeek-V4.1-Flash-1017 is the Flash checkpoint DeepSeek serves on the website, the app, and the API, and the model card sets that checkpoint beside Claude Opus 5 and GPT-5.6 Sol.

The split is list price and call surface. DeepSWE and Terminal-Bench say how a lab scored a checkpoint on coding-agent work. Price and call surface say which account a buyer can keep open for a month of real use.

DeepSeek's card prints 77.0 on DeepSWE and 42.3 on Terminal-Bench.

DeepSWE is DeepSeek's software-engineering eval. Terminal-Bench is a public suite of terminal tasks, and the task code lives in the public terminal-bench repository (terminal-bench on GitHub). The Hugging Face card for this checkpoint prints 77.0 on DeepSWE and 42.3 on Terminal-Bench (Flash-1017 model card). MarkTechPost published a same-week account of the release (MarkTechPost).

Those two figures are tied to the October 17, 2026 card.

A software-engineering eval asks a model to change code until tests pass or a stated bug is gone. A terminal eval asks it to work in a shell: create files, run commands, and meet a checker at the end. DeepSWE and Terminal-Bench sit in that family. They are built for agent loops, tool use, and multi-step jobs, which is a different exam from a short chat reply about a concept.

Label the suite in the same breath as the number. DeepSWE and SWE-bench are different suites, so a DeepSWE figure should stay labeled DeepSWE when you repeat it. Mixing the names puts a 77.0 on a different test from the one DeepSeek scored.

The 77.0 and the 42.3 do not, by themselves, establish a shared-run win over Opus 5 or Sol.

For a coding buyer, the shape of the exam still carries information. If your week is pull requests, failing tests, and a terminal session, you are looking at the same kind of work these suites stage. If your week is email, a policy draft, or a pile of PDFs, a terminal score is a weak guide to the product you should open. The card is evidence about agent-shaped coding. It is thin evidence about ordinary office chat.

That is the scope I would give these two numbers.

Research work wants a model that can hold a source and answer from it. DeepSWE and Terminal-Bench stage code and shells. A literature summary is a separate job, and I would choose the chat product whose sources view you already like for that job. The suite names tell you the exam was a repo and a shell. Your pile of PDFs is a different desk.

The suffix 1017 is the checkpoint ID. Earlier Flash writeups refer to earlier builds, and the scores in those writeups belong to those builds. Keep 1017 in the sentence whenever you quote 77.0 or 42.3. A bare "DeepSeek Flash" line will mix this card with a previous one, and the two cards are allowed to disagree.

When you paste these results into a team doc, put the model ID, the suite, the score, and the source URL on one line. A line that reads Flash-1017, DeepSWE, 77.0, model card is a measurement. A line that reads DeepSeek beats Opus is a slogan. The same rule applies to 42.3. Write Flash-1017, Terminal-Bench, 42.3, model card. Then write a second line for price that points at each vendor's pricing page, because the score line and the price line answer different questions. If someone later swaps in a newer Flash checkpoint, the dated ID forces them to fetch a new card before they reuse 77.0. That small habit prevents yesterday's table from running tomorrow's agent budget. I would reject a comparison slide that shows three logos and one number with the suite name cropped off.

77.0 and 42.3 are card scores for one checkpoint. The account you open still depends on the invoice and the app in front of you.

List price and the place you call the model separate these three.

OpenAI's API pricing page lists GPT-5.6 Sol at $4 per million input tokens and $20 per million output tokens (OpenAI API pricing). Anthropic's launch pricing put Claude Opus 5 at half of Fable 5's API price (Anthropic pricing).

Agent runs are long. A model that writes a plan, calls tools, reads the tool output, and tries again will emit far more output tokens than a one-paragraph answer. The weekly bill follows that shape. A subscription chat plan hides the meter inside a flat fee until you hit the plan cap. An API agent prints the meter on every step. Those are different purchases even when the benchmark names on the slide match.

Hosting is the other half of call surface. Opus 5 and Sol are reached through their vendors' APIs and chat products. Flash-1017 is the checkpoint with a public model card on Hugging Face, so a team that must read weights, license terms, and serving notes has a document to open. A team that will only ever type into a chat box can ignore the file repo and stay in the app they already pay for.

Self-hosting is a project of its own. You bring GPUs, a runtime, monitoring, and a person who can patch the stack. That cost sits outside the token rate. For a solo user, the hosted app or the hosted API is the practical path, and the token rate is the number that belongs on the spreadsheet. Renting a GPU box to save a chat subscription is a second project, and it deserves its own approval.

A company buy has a procurement shape a benchmark row never captures. Security review, data-processing terms, and whether the vendor will sign the company's paper all sit in the deal. Opus and Sol are sold as enterprise APIs with an account team around them. DeepSeek's API is a developer account you can start from the docs, which some security teams will accept and some will send back for a review cycle. I would ask the security owner which of those paths is already approved before I argued about DeepSWE in the same meeting.

Read the rows as four different decisions. Suite scores stay on the Flash card's line. Price stays on the vendor page for each API. Call surface stays on the product you can open this week. A single winner across those rows would require one buyer with one job, and that buyer is rare.

The right account follows the task you actually run.

Opus 5 is the Claude flagship in this matchup, and Sol is the GPT-5.6 tier OpenAI sells on the main API. Flash-1017 is the new checkpoint, so it carries the freshest card and the least outside commentary. Freshness is an advantage for anyone who wants the current Flash coding line. It is a reason for a cautious buyer to keep the flagship chat account in place while those two scores sit on one lab's card. Budget the experiment in API calls you can afford to throw away. Keep the production writing in the account your team already knows how to support.

I would put a non-technical person on Sol or Opus for everyday writing, reading, and short research. Those products are built as chat apps, and the bill is the subscription they already understand. Flash earns the seat when the work is a repository, a shell, or an agent that will run many steps, and when the buyer is willing to use DeepSeek's app or API.

Split it by job.

• Office research and drafts. Sol and Opus. The chat products are the whole interface. You paste a source, ask for a summary, and stay inside one account.
• Coding agents and terminal work. Flash-1017 is the checkpoint whose public card speaks DeepSWE and Terminal-Bench. That is the evidence attached to this release. Use it when the job looks like those evals.
• A mixed week. Keep a frontier chat subscription for ambiguous writing, and point the agent stack at Flash when the stack is the thing you are building.

Three starting points cover the buyers who search this matchup. A person who writes, reads, and summarizes all day should open Sol or Opus and stay in the chat app. The work is language, the interface is already built, and a terminal score will steer that account poorly. A developer connecting an agent to a repository should put Flash-1017 on the shortlist, then read DeepSeek's pricing page before any loop runs overnight. The card is the reason it makes the shortlist. The pricing page is the reason it stays. A team lead picking a house model should standardize on the vendor the company has already approved, then add a second API for the job the first one bills harshly. Agent volume is the usual candidate for that second API. Production prose can stay where support already exists.

Until a second lab posts Terminal-Bench beside the October 17, 2026 card, buyers should price the three APIs from the vendor pages and treat 42.3 as Flash-1017's published Terminal-Bench score. The October 17, 2026 card is the Terminal-Bench source behind that 42.3.