Three AI labs shipped flagship models within five weeks of each other this spring, and the gap between them has little to do with raw intelligence anymore. Anthropic’s Claude Opus 4.8 landed on May 28, 2026 with the highest SWE-Bench Verified score of any released model at the time. DeepSeek followed on April 24 with V4 Pro, then cut its API price by 75% a month later. MiniMax, a Shanghai lab better known outside China for video generation than coding agents, closed the window on June 1 with M3, an open-weight model priced at a fraction of what Anthropic charges per token.
The result is a genuine three-way tradeoff rather than a single winner. Claude Opus 4.8 wins on capability. DeepSeek V4 Pro wins on raw cost efficiency at the API level. MiniMax M3 wins on openness, self-hosting flexibility, and native multimodal input. Picking between them now depends less on “which model is smartest” and more on what a team actually optimizes for: benchmark ceiling, dollar-per-task, or the ability to run the weights on its own hardware.
This comparison breaks down the specs, benchmark scores from independent leaderboards, real API pricing, and which model fits which team, using only data published between April and July 2026.
It also matters who is asking. A solo developer choosing between these three cares about a different set of numbers than a platform team signing an annual enterprise contract. The specs below are the same for both readers. What each reader should do with them is not, and this piece tries to separate the two rather than flatten them into one generic recommendation.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
MiniMax M3 vs Claude Opus 4.8 vs DeepSeek V4 Pro: The Three Models at a Glance
Before the deep dive, here is the short version. Each of these three models shipped in a six-week window between late April and early June 2026, and each one is optimized for a different priority.
| Model | Maker | Released | Weight Status | Price In/Out (per 1M tokens) | Context Window | Best Known For |
|---|---|---|---|---|---|---|
| MiniMax M3 | MiniMax | June 1, 2026 | Open-weight | $0.60 / $2.40 | 1,000,000 tokens | Cheapest frontier-adjacent coding and multimodal agent |
| Claude Opus 4.8 | Anthropic | May 28, 2026 | Closed | $5.00 / $25.00 | 200,000 tokens | Highest SWE-Bench Verified score of any released model |
| DeepSeek V4 Pro | DeepSeek | April 24, 2026 | Closed | $0.435 / $0.87 | 1,000,000 tokens (384K max output) | Lowest cost per token of the three after its May price cut |
MiniMax M3 is the newcomer to this particular matchup. It is a Mixture-of-Experts model with 428 billion total parameters and 23 billion active parameters per token, built around a new attention mechanism MiniMax calls MSA, or MiniMax Sparse Attention, according to the official MiniMax M3 announcement. Claude Opus 4.8 is Anthropic’s current flagship, positioned for agentic coding and computer-use tasks where accuracy matters more than price. DeepSeek V4 Pro sits in between: closed like Opus 4.8, but priced closer to MiniMax M3 after DeepSeek slashed rates in May.
None of these three models had been benchmarked against the other two in a single article before this one, which is part of why the comparison matters. Most coverage so far has pitted MiniMax M3 against DeepSeek V4 Pro alone, or Opus 4.8 against OpenAI and Google’s flagships. Putting all three side by side answers a more practical question: for a given task and budget, which one should you actually call?
Full Specs Comparison: MiniMax M3, Claude Opus 4.8, and DeepSeek V4 Pro
Specs alone do not tell you which model to pick, but they set the boundaries of what each one can do. Here is the full breakdown across fourteen data points.
| Spec | MiniMax M3 | Claude Opus 4.8 | DeepSeek V4 Pro |
|---|---|---|---|
| Developer | MiniMax | Anthropic | DeepSeek |
| Release date | June 1, 2026 | May 28, 2026 | April 24, 2026 |
| Weight availability | Open-weight | Closed | Closed |
| Architecture | Mixture-of-Experts with MiniMax Sparse Attention | Not publicly disclosed | Not publicly disclosed |
| Parameter count | 428B total / 23B active | Not disclosed | Not officially disclosed |
| Context window | 1,000,000 tokens | 200,000 tokens | 1,000,000 tokens |
| Max output | Not specified | Not specified | 384,000 tokens |
| Input price (per 1M tokens) | $0.60 | $5.00 | $0.435 |
| Output price (per 1M tokens) | $2.40 | $25.00 | $0.87 |
| SWE-Bench Verified | 80.5% | 88.6% | 73.6% |
| SWE-Bench Pro | 59.0% | Not published | 55.4% |
| Multimodal input | Text, image, video, computer use | Text, vision, computer use | Text only |
| Self-hostable | Yes, via Hugging Face weights | No | No |
| API access points | MiniMax API, Hugging Face, OpenRouter, Fireworks AI, EvoLink | Anthropic API, OpenRouter | DeepSeek API, OpenRouter, DeepInfra |
Two things jump out immediately. First, both MiniMax M3 and DeepSeek V4 Pro ship a 1 million token context window, five times larger than Claude Opus 4.8’s 200,000 tokens. Second, the price gap is not small. Opus 4.8’s input pricing runs about 8.3 times higher than MiniMax M3 and roughly 11.5 times higher than DeepSeek V4 Pro. That gap gets even wider on output tokens, where Opus 4.8 charges 10.4 times more than MiniMax M3 and 28.7 times more than DeepSeek V4 Pro.
Neither DeepSeek nor Anthropic has published exact parameter counts for its model, and MiniMax is the only one of the three that discloses architecture details at all. That is a direct consequence of licensing. Open-weight models tend to publish more of their internals because researchers can inspect the weights directly anyway. Closed models keep that information proprietary regardless of price.
Release Dates and Why Five Weeks Changed the Market
The timing here is not a coincidence. DeepSeek V4 Pro shipped first, on April 24, 2026, at an initial list price of $1.74 per million input tokens and $3.48 per million output tokens. A month later, DeepSeek made a permanent 75% cut to that pricing, bringing it down to the current $0.72.435 and $0.87 rates. That move put immediate pressure on every other lab selling API access to a frontier-class model.
Anthropic released Claude Opus 4.8 on May 28, 2026, four days after DeepSeek’s price cut took hold. Opus 4.8 did not compete on price. Instead, it pushed the capability ceiling higher, landing an Artificial Analysis Intelligence Index score of 61.4 and a SWE-Bench Verified score of 88.6%, both leading figures among released models at the time, according to Artificial Analysis’s model comparison data. Anthropic’s bet was that enterprise buyers would pay a premium for the model that makes the fewest mistakes on agentic coding tasks, rather than the cheapest one.
MiniMax M3 arrived four days after that, on June 1, and took a third path entirely. Rather than compete purely on price like DeepSeek or purely on capability like Anthropic, MiniMax released the model’s weights publicly through Hugging Face, live within a week of the announcement. That let any team with enough GPU capacity self-host a model that scores 80.5% on SWE-Bench Verified, within roughly 8 points of Opus 4.8’s score, at a fraction of the ongoing API cost. MiniMax also priced its own hosted API at $0.60 per million input tokens, undercutting Opus 4.8 by more than 8 times.
The net effect: three different strategic bets landed inside a five-week window, and none of the three labs blinked. That is unusual. In past product cycles, a price cut from one vendor often triggered matching cuts from the others within days. This time, Anthropic held its pricing steady and leaned entirely on capability, while MiniMax used openness itself as a pricing lever, since self-hosted inference carries no per-token API cost once the infrastructure is running.
Benchmark Performance: Coding, Reasoning, and Agentic Tasks
Benchmark scores for all three models come from a mix of vendor release notes and independent leaderboards, and it is worth being explicit about which is which. MiniMax’s SWE-Bench Pro figure, for example, is vendor-reported. DeepSeek and Anthropic’s SWE-Bench Verified numbers appear on third-party trackers including LLM-Stats.com and Artificial Analysis, which independently run the evaluation suite rather than relying on lab press releases.
SWE-Bench Verified and SWE-Bench Pro Scores
SWE-Bench Verified measures how often a model can resolve real GitHub issues pulled from popular open-source repositories, with a human-verified solution to check against. It is the closest thing the industry has to a standard for “can this model actually fix bugs.” Claude Opus 4.8 leads all three at 88.6%. MiniMax M3 follows at 80.5%, an 8.1 point gap. DeepSeek V4 Pro trails at 80.6%, an 8 point gap from Opus 4.8.8.
SWE-Bench Pro is a harder, newer variant built around production-grade repositories rather than curated open-source issues, and Anthropic has not published an Opus 4.8 score on it as of this writing. Between the two models that have reported scores, MiniMax M3 edges out DeepSeek V4 Pro, 59.0% to 55.4%.
Reasoning and Long-Horizon Benchmarks
On reasoning-heavy tasks, the picture shifts again. DeepSeek V4 Pro posts a GPQA Diamond score of 90.1%, a graduate-level science reasoning benchmark, while Artificial Analysis lists MiniMax M3 with an Intelligence Index score of 44 among open-weight models, well below Opus 4.8’s 61.4. On BrowseComp, a benchmark for web research and multi-step retrieval tasks, MiniMax M3 and DeepSeek V4 Pro are close, at 83.5% and 83.4% respectively.
| Benchmark | MiniMax M3 | Claude Opus 4.8 | DeepSeek V4 Pro |
|---|---|---|---|
| SWE-Bench Verified | 80.5% | 88.6% | 73.6% |
| SWE-Bench Pro | 59.0% | Not published | 55.4% |
| Terminal-Bench 2.1 | 66.0% | Not published | 67.9% |
| GPQA Diamond | Not published at this tier | Not published in this dataset | 90.1% |
| BrowseComp | 83.5% | Not published | 83.4% |
| LM Arena Elo (text) | Not ranked in this dataset | 1,486 | Not ranked in this dataset |
| Artificial Analysis Intelligence Index | 44 (open-weight tier) | 61.4 | Not published (non-reasoning variant) |
Read together, the benchmark data tells a consistent story. Opus 4.8 leads on the benchmarks that measure real-world coding accuracy and general reasoning breadth. DeepSeek V4 Pro is competitive on structured reasoning and slightly ahead on Terminal-Bench 2.1, a benchmark for command-line and agentic terminal tasks. MiniMax M3 sits close behind both on coding-specific tasks while trailing on the broader Intelligence Index, which aggregates multiple reasoning evaluations rather than just coding.
Agentic Browsing, Computer Use, and Inference Speed
Two additional benchmarks show up specifically in MiniMax’s own release data and are worth flagging separately, since neither DeepSeek nor Anthropic has published directly comparable scores. MiniMax reports M3 at 70.0% on OSWorld-Verified, a benchmark that scores how reliably a model completes multi-step tasks inside a real desktop operating system, and 74.2% on MCP Atlas, which measures tool-use accuracy through the Model Context Protocol. Both are vendor-reported figures rather than independently audited, so they should be read as MiniMax’s own claim about its computer-use and tool-calling ability rather than a neutral third-party result.
MiniMax also claims a substantial inference speed advantage at long context lengths: up to 15 times faster decoding compared to its own prior M2.7 release, plus a 9.4 times speedup from a custom CUDA FP8 kernel MiniMax built for Hopper-generation GPUs. Neither DeepSeek nor Anthropic has published a directly comparable throughput figure for V4 Pro or Opus 4.8, so this is another spec where MiniMax is simply more forthcoming with internal engineering detail, consistent with its open-weight approach.
API Pricing Compared: Cost Per Million Tokens
Pricing is where this comparison gets interesting, because the spread between these three models is larger than almost any other three-way frontier comparison this year.
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Notes |
|---|---|---|---|
| MiniMax M3 | $0.60 | $2.40 | Launch promotion briefly dropped this to $0.30 / $1.20 in the first week |
| Claude Opus 4.8 | $5.00 | $25.00 | No public promotional discount reported as of this writing |
| DeepSeek V4 Pro | $0.435 | $0.87 | Down from $1.74 / $3.48 after a permanent 75% cut in May 2026. Cache-hit input is priced separately at $0.003625/1M |
Put in concrete terms: a workload that processes 10 million input tokens and 2 million output tokens in a month costs roughly $100 on Claude Opus 4.8 ($50 for input, $50 for output). The same workload runs about $10.80 on MiniMax M3, and about $6.09 on DeepSeek V4 Pro. That is not a rounding difference. It is the gap between a project that needs budget approval and one that a single engineer can expense without asking anyone.
The catch is that price alone does not capture task success rate. If DeepSeek V4 Pro or MiniMax M3 needs two or three retries to match what Opus 4.8 gets right on the first pass, the effective cost gap shrinks once wasted tokens and engineering time spent debugging failed outputs enter the calculation. For high-stakes agentic workflows, the 15-point SWE-Bench Verified gap between DeepSeek V4 Pro and Opus 4.8 can matter more than the 11.5x sticker price difference.
Open-Weight vs Closed-Weight: What You’re Actually Buying
MiniMax M3 is the only model in this comparison you can download and run yourself. Its weights are published on Hugging Face, which means a team with sufficient GPU capacity can self-host the model with no per-token API fee at all, no rate limits beyond what the hardware allows, and no dependency on MiniMax’s uptime. That comes with tradeoffs: someone has to provision and maintain the GPUs, patch the inference stack, and handle scaling, none of which is free even if the token cost is zero.
Claude Opus 4.8 and DeepSeek V4 Pro are both closed. You access them exclusively through hosted APIs, which means Anthropic and DeepSeek control availability, rate limits, and any future price changes. For DeepSeek V4 Pro, that also means the underlying architecture and exact parameter count are not publicly documented, beyond an unconfirmed third-party estimate that DeepSeek itself has not verified.
The practical difference shows up in three places: data residency, since self-hosting keeps prompts and outputs entirely on infrastructure a team controls, long-term cost at scale, since a self-hosted model’s marginal cost per request drops toward the cost of electricity and hardware amortization once volume is high enough, and vendor risk, since an API-only model can change its pricing or availability terms with notice, while a downloaded set of weights keeps working even if the vendor disappears tomorrow.
There is a middle path worth mentioning too. A team does not have to choose between fully self-hosted and fully API-dependent on day one. Plenty of engineering organizations start on MiniMax’s hosted API to validate that the model fits their task, then move to self-hosting later once volume justifies the GPU spend. That staged approach is only possible because the weights are public. Neither Claude Opus 4.8 nor DeepSeek V4 Pro offers that off-ramp, since there is no version of either model a customer can eventually take in-house no matter how much API volume they run.
Context Windows and Long-Context Handling
MiniMax M3 and DeepSeek V4 Pro both ship a 1 million token context window, roughly five times larger than Claude Opus 4.8’s 200,000 tokens. For teams working with large codebases, long legal contracts, or entire documentation sets in a single prompt, that difference is not academic. A 1 million token window can hold an entire mid-sized monorepo. A 200,000 token window generally cannot.
DeepSeek V4 Pro caps output at 384,000 tokens even though its input window reaches a full million, according to DeepSeek’s official pricing and API documentation. MiniMax has not published a separate output cap distinct from its context window in the sources reviewed for this piece. Anthropic has not published an expanded context option for Opus 4.8 at the time of writing, keeping it at 200,000 tokens across the board.
Context window size does not automatically translate into usable long-context performance. A model can accept a million tokens and still lose track of details buried in the middle of that input, a failure mode researchers often call the “lost in the middle” problem. None of the three vendors in this comparison has published independent long-context retrieval scores specific enough to settle which model handles a genuinely full million-token prompt best, so this is one area where a team’s own testing matters more than any published spec.
Multimodal Capabilities: Text, Image, and Video
MiniMax M3 is the most multimodal of the three by a clear margin. It accepts text, image, and video input natively, and MiniMax’s own release materials describe a computer-use capability that lets the model operate a desktop environment directly, per the MiniMax M3 launch post. That combination, native video input alongside computer use, is not something either of the other two models in this comparison currently advertises.
Claude Opus 4.8 supports vision input, and Anthropic has built out agentic computer-use capabilities in prior Opus releases, positioning the model for tasks like navigating a browser or filling out a form autonomously. DeepSeek V4 Pro, by contrast, is confirmed as text-only across the sources reviewed for this article. Its API documentation does not list image, video, or audio input support, which narrows its use case to text and code-heavy workloads rather than anything involving visual understanding.
For teams building agents that need to read a screenshot, parse a video clip, or click through a web interface, that gap rules DeepSeek V4 Pro out entirely, regardless of how attractive its pricing looks on paper.
Platform and API Availability
All three models are reachable through more than one channel, which matters for teams that want to avoid hard-coding a single vendor’s SDK into their stack. MiniMax M3 is available directly through MiniMax’s own API, as self-hosted weights via Hugging Face, and through third-party inference providers including OpenRouter and Fireworks AI, plus EvoLink.
Claude Opus 4.8 is available through Anthropic’s own API and is also listed on OpenRouter, alongside Anthropic’s published pricing page. DeepSeek V4 Pro is available through DeepSeek’s own API, through OpenRouter, and through DeepInfra.
The fact that all three now sit on OpenRouter’s unified endpoint is one of the more useful developments buried in this comparison. It means a team can test all three models against the same evaluation harness, using the same request format, by changing a single model identifier string rather than rewriting integration code three separate times.
The choice of access point changes the economics too, not just the convenience. Going direct to MiniMax, Anthropic, or DeepSeek’s own API usually gets the lowest published price and first access to new features, since the lab controls that channel directly. Routing through OpenRouter adds a small margin on top of the base price in exchange for a single API key across dozens of models and automatic failover if one provider has an outage. Fireworks AI and DeepInfra sit somewhere in between: both run their own inference infrastructure for open models like MiniMax M3, which can mean different latency and throughput characteristics than MiniMax’s own hosted API, even at similar pricing. For a team still evaluating which model to commit to, that flexibility is often worth more than the small margin difference.
Real-World Use Cases: Five Scenarios Where Each Model Wins
Specs and benchmarks matter, but most teams choose a model based on a specific job to be done. Here are five concrete scenarios pulled from how these three models are actually being positioned in the market.
- A seed-stage startup building an autonomous coding agent on a fixed runway. MiniMax M3’s $0.60 / $2.40 pricing and 80.5% SWE-Bench Verified score make it the practical choice when every dollar of burn rate counts and the team cannot justify Opus 4.8’s per-token premium.
- An enterprise team automating code review across a regulated financial services codebase. Claude Opus 4.8’s 88.6% SWE-Bench Verified score and stronger track record on agentic accuracy justify the higher API cost when a bad autonomous merge could cause real damage.
- A customer support platform running millions of reasoning calls a day. DeepSeek V4 Pro’s $0.87 output price, the lowest of the three, makes the per-ticket economics work at a volume where even small per-call savings compound into real budget impact.
- A research lab that needs to fine-tune or inspect model internals for a published paper. MiniMax M3 is the only one of the three whose weights are publicly downloadable, making it the only option for teams that need to modify the model itself rather than just call an API.
- A legal-tech company processing entire contract sets or discovery documents in one pass. MiniMax M3 and DeepSeek V4 Pro’s 1 million token context windows can hold documents that would need to be chunked and split across multiple calls on Claude Opus 4.8’s 200,000 token limit.
A sixth scenario is worth calling out separately: any product that needs to process video or let a model operate a desktop UI directly is currently a MiniMax M3-only conversation among these three, since neither Opus 4.8 nor DeepSeek V4 Pro currently offers confirmed native video input.
Migration Guide: Switching Your Stack Between These Models
Moving a production workload from one of these models to another is rarely a one-line change, even when the API is nominally compatible. Here is a practical sequence for teams testing a switch.
- Build an evaluation set from your own traffic. Public benchmark scores tell you how a model performs on SWE-Bench or GPQA Diamond, not on your specific prompts. Pull 50 to 100 real requests and their known-good outputs before changing anything.
- Check context and output limits against your actual payloads. If you are moving from MiniMax M3 or DeepSeek V4 Pro’s 1 million token window down to Opus 4.8’s 200,000 tokens, confirm your longest real prompts still fit, or plan a chunking strategy.
- Re-test system prompt conventions. Each lab tunes its model against slightly different system prompt patterns. A prompt that reliably steers Opus 4.8 may need rewording to get equivalent behavior from DeepSeek V4 Pro or MiniMax M3.
- Verify tool-calling and function schema compatibility. Agentic workflows that depend on structured tool calls should be tested explicitly, since exact JSON schema handling can differ even across otherwise OpenAI-compatible endpoints.
- Route through a unified endpoint during testing. Since all three models are reachable through OpenRouter, testing can happen with a single integration rather than three.
- Run a shadow or canary rollout before cutting over fully. Send a percentage of live traffic to the new model and compare output quality and cost side by side before retiring the old integration.
Because MiniMax M3, DeepSeek V4 Pro, and Claude Opus 4.8 are all listed on OpenRouter, a basic model swap in code can look as simple as changing one string, as shown below.
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_KEY",
)
# Swap between models by changing only this string
for model_id in [
"minimaxi/minimax-m3",
"anthropic/claude-opus-4.8",
"deepseek/deepseek-v4-pro",
]:
response = client.chat.completions.create(
model=model_id,
messages=[{"role": "user", "content": "Summarize this pull request diff."}],
)
print(model_id, "->", response.choices[0].message.content[:200])
The code stays identical across all three. What changes underneath is cost, latency, output quality, and how much context you can safely send, which is exactly why the evaluation step matters more than the integration step.
Pros and Cons: MiniMax M3, Claude Opus 4.8, and DeepSeek V4 Pro
MiniMax M3 pros: open-weight and self-hostable, cheapest hosted API of the three on input tokens, native image and video input, computer-use capability, 1 million token context window, available on five separate platforms.
MiniMax M3 cons: trails Opus 4.8 by 8.1 points on SWE-Bench Verified, lowest Artificial Analysis Intelligence Index of the three, self-hosting requires real GPU infrastructure to realize the cost benefit, and some benchmark figures are vendor-reported rather than independently verified.
Claude Opus 4.8 pros: highest SWE-Bench Verified score of the three at 88.6%, highest Artificial Analysis Intelligence Index at 61.4, strong agentic and computer-use track record, backed by Anthropic’s enterprise contractual terms and support.
Claude Opus 4.8 cons: most expensive of the three by a wide margin, smallest context window at 200,000 tokens, closed weights with no self-hosting option, no public promotional pricing found at the time of writing.
DeepSeek V4 Pro pros: lowest output price of the three at $0.87 per million tokens, strong GPQA Diamond score at 90.1%, 1 million token context window, available through three separate access points including OpenRouter.
DeepSeek V4 Pro cons: text-only, no confirmed image, video, or audio input, closed weights, trails both other models on SWE-Bench Verified at 73.6%, parameter count and architecture remain undisclosed.
Which Model Should You Choose
The right pick depends on which constraint binds hardest for your team: budget, benchmark ceiling, data control, or input modality. The table below maps common team profiles to a recommendation.
| Team Profile | Recommended Model | Why |
|---|---|---|
| Budget-constrained startup shipping an agentic coding tool | MiniMax M3 | Cheapest frontier-adjacent pricing plus open weights as a self-hosting fallback |
| Enterprise needing the top coding benchmark score and computer-use agents | Claude Opus 4.8 | Highest SWE-Bench Verified score (88.6%) among the three |
| High-volume support or reasoning pipeline sensitive to output-token cost | DeepSeek V4 Pro | Lowest output price at $0.87 per million tokens after the May cut |
| Team needing data sovereignty or on-prem deployment | MiniMax M3 | Only one of the three with publicly downloadable weights |
| Long-document or codebase analysis, contracts, monorepos | MiniMax M3 or DeepSeek V4 Pro | Both offer a 1 million token context window versus Opus 4.8’s 200,000 |
| Multimodal agent needing video or image input | MiniMax M3 | Only model of the three with confirmed native image and video input |
| Regulated industry cautious about routing data through a Chinese vendor’s API | Claude Opus 4.8 | US-based, closed model with established enterprise contract terms |
MiniMax M3 shows up as the recommendation in four of seven rows above. That is less an endorsement of one model over the others and more a reflection of how much weight open licensing and multimodal input carry once cost enters the equation. For teams that specifically need the highest possible accuracy on agentic coding tasks and can absorb the price, Opus 4.8 still leads outright.
The Verdict: Cost, Capability, and the Right Call for June 2026
There is no single winner across all three models, and pretending otherwise would misrepresent the data. Claude Opus 4.8 is the most capable model in this group by every coding and reasoning benchmark reviewed here, leading MiniMax M3 by 8.1 points and DeepSeek V4 Pro by 15 points on SWE-Bench Verified. If a workload cannot tolerate mistakes, Opus 4.8 earns its premium.
DeepSeek V4 Pro wins on pure cost efficiency, especially on output tokens, where it charges roughly 3.6 cents for every dollar Opus 4.8 charges. That makes it the strongest option for high-volume, lower-stakes reasoning workloads where an occasional error costs less than the API bill would otherwise run.
MiniMax M3 threads the needle between the two. It is not the cheapest API on paper (DeepSeek V4 Pro undercuts it on both input and output pricing), and it is not the most capable (Opus 4.8 beats it on every published coding benchmark). What MiniMax M3 offers instead is optionality: open weights that remove vendor dependency entirely, native video and image input that neither competitor currently matches, and a 1 million token context window at roughly a tenth of Opus 4.8’s price. For teams that value flexibility over squeezing out the last percentage point of benchmark score or the last cent of API cost, that combination is hard to find anywhere else released in the same six-week window.
The practical takeaway: run your own evaluation set against all three before committing, since the 8 to 15 point benchmark gaps documented here can shrink or widen depending on your specific task, and the pricing gap is large enough that even a partial migration to a cheaper model can meaningfully change a project’s monthly bill.
One more point worth stating plainly: none of these three rankings are permanent. DeepSeek cut its price by 75% a month after launch, MiniMax undercut Opus 4.8’s pricing by more than 8 times on day one, and Anthropic has a history of shipping follow-on releases within months rather than years. A comparison this specific has a shelf life. Treat the benchmark scores and pricing here as a snapshot of June and July 2026, confirm the current numbers directly with each vendor before signing any contract or committing production infrastructure, and expect at least one of these three models to change its pricing or publish a successor before the end of the year.
For the full ranked field of 2026 releases, see the best AI models 2026 hub.
Frequently Asked Questions
Is MiniMax M3 really free to self-host?
The weights are free to download from Hugging Face, but self-hosting is not free in practice. A team still needs enough GPU capacity to run a 428 billion parameter Mixture-of-Experts model, which for most organizations means renting cloud GPU instances or building out on-prem hardware. The zero cost comes from the licensing, not from the infrastructure bill.
Which model has the best benchmark scores for coding specifically?
Claude Opus 4.8 leads on SWE-Bench Verified at 88.6%, ahead of MiniMax M3 at 80.5% and DeepSeek V4 Pro at 73.6%. On the newer SWE-Bench Pro benchmark, where Opus 4.8 has not published a score, MiniMax M3 leads DeepSeek V4 Pro, 59.0% to 55.4%.
How much does it cost to run these models at scale?
For a workload of 10 million input tokens and 2 million output tokens a month, expect roughly $100 on Claude Opus 4.8, $10.80 on MiniMax M3, and $6.09 on DeepSeek V4 Pro, based on published per-token pricing as of July 2026.
Can any of these models handle image or video input?
MiniMax M3 supports native text, image, and video input, plus a computer-use mode. Claude Opus 4.8 supports text and vision input along with agentic computer-use capability. DeepSeek V4 Pro is confirmed text-only across the sources reviewed for this article.
Do I need a 1 million token context window, or is 200,000 tokens enough?
It depends on what is being fed to the model. A 200,000 token window, Claude Opus 4.8’s limit, comfortably fits most single documents, chat histories, and small-to-medium codebases. A full 1 million token window, available on MiniMax M3 and DeepSeek V4 Pro, matters most for entire monorepos, large discovery document sets, or very long multi-document research tasks.
Is DeepSeek V4 Pro a safe choice for enterprise use given it is a Chinese company?
That depends entirely on an organization’s own data governance policy and any regulatory requirements it operates under. Some regulated industries and government-adjacent teams explicitly avoid routing sensitive data through vendors headquartered outside their home jurisdiction, which is one reason Claude Opus 4.8, as a US-based closed model with enterprise contract terms, appears as the recommended pick for that use case in this comparison. This article does not make a legal or compliance determination, so consult your own counsel or security team.
Which model is best for agentic, tool-calling workflows?
Claude Opus 4.8 has the strongest published track record on agentic and computer-use tasks among the three, backed by its leading SWE-Bench Verified score. MiniMax M3 is a close second on coding-specific agentic benchmarks like SWE-Bench Pro and includes its own computer-use mode. DeepSeek V4 Pro scores competitively on Terminal-Bench 2.1 but has not published dedicated computer-use capability in the sources reviewed here.
What is the easiest way to test all three models before committing to one?
Since MiniMax M3, Claude Opus 4.8, and DeepSeek V4 Pro are all listed on OpenRouter’s unified API, a team can test all three with the same integration code by changing only the model identifier string, as shown in the migration guide code sample earlier in this article. Build a small evaluation set from real prompts first, since published benchmark scores may not reflect performance on any specific task.
