Three small, cheap models shipped within weeks of each other, and the price gap between them will settle more buying decisions than the benchmark gap. Anthropic released Claude Haiku 5.5 earlier this month. DeepSeek pushed its V4.1-Flash-1017 revision on October 17. OpenAI introduced GPT-6.1 Terra on October 26, one day before this article. On the list prices OpenAI and Anthropic publish, Terra costs $1 per million input tokens and Haiku 5.5 costs $0.10, a 10x difference before caching or batch discounts enter the picture.
This comparison puts the three side by side for developers who run high-volume work: classification, summarization, support bots, browser agents and coding subagents. It covers DeepSeek API pricing next to the two Western rivals, the cost of four realistic workloads, what each vendor has and has not published on benchmarks, and a step-by-step migration path. Where a number is not confirmed on an official page, the article says so and leaves it out of the tables.
One note on scope. The site already ran a broader matchup of DeepSeek, Haiku 5.5 and Gemini Flash, which you can read in our earlier DeepSeek, Haiku and Gemini Flash comparison. This piece is different. It centers on the newest checkpoint of each line (Terra is a day old, and the 1017 revision of DeepSeek arrived after the September V4.1 Flash release) and on what the bills look like at scale.
Start with who each model is for. Claude Haiku 5.5 is Anthropic’s small model, and the company calls it “the cheapest, fastest, and most capable small model we’ve ever released.” Anthropic aims it at summarization, subagents, browser use and live customer support. GPT-6.1 Terra is the newest model in OpenAI’s GPT-6 series, tuned for fast, efficient, high-volume professional work, image analysis and computer use. DeepSeek V4.1 Flash 1017 is the October 17 revision of the V4.1 Flash model that DeepSeek released on September 10, and it ships as open weights as well as through a hosted API.
The three sit in the same budget tier but make different trades. Haiku 5.5 wins on published price. Terra offers a cheap cached-input rate and the weight of OpenAI’s tooling in ChatGPT Work and Codex. DeepSeek is the only one of the three you can download and run yourself, which changes the cost math for teams with spare GPUs and strict data rules. None of them is a frontier flagship. If you need maximum reasoning, look at the Opus, Sol and Astra tier instead, and expect to pay a multiple of what follows.
For readers who want the short version before the detail: Haiku 5.5 is the cheapest verified option, Terra is the strongest pick inside the OpenAI stack, and DeepSeek is the pick when open weights or the lowest reported rates matter more than a clean, confirmed price sheet.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Haiku 5.5, GPT-6.1 Terra and DeepSeek V4.1 Flash 1017: Specs Table
The table below uses only figures published on the vendors’ own pages or confirmed across multiple sources on October 27, 2026. Where DeepSeek’s per-token rates would go, the table says the rate sheet is unconfirmed rather than guessing. Third-party trackers disagree with each other on those rates, and the section on pricing explains how.
| Spec | Claude Haiku 5.5 | GPT-6.1 Terra | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|
| Vendor | Anthropic | OpenAI | DeepSeek |
| Newest release | October 7, 2026 | October 26, 2026 | October 17, 2026 (revision of the September 10 V4.1 Flash) |
| API identifier | claude-haiku-5-5 | gpt-6.1-terra | deepseek-flash (stable alias) |
| Official positioning | Fastest, most capable Haiku. High-volume, cost-sensitive work | Newest GPT-6 model. Fast, efficient, high-volume professional work, image analysis, computer use | Flash tier, native multimodal API since September 10, agentic coding focus |
| Input price per 1M tokens | $0.10 (prompts up to 100,000 tokens) | $1.00 (Standard) | Unchanged by the 1017 update (rate sheet not confirmed in our sources) |
| Output price per 1M tokens | $0.50 (prompts up to 100,000 tokens) | $5.00 (Standard) | Unchanged by the 1017 update (rate sheet not confirmed in our sources) |
| Long-prompt pricing | $0.50 input and $2.50 output above 100,000 tokens | Cache-write prices are “before long-context adjustments” on OpenAI’s page | Not confirmed |
| Cached input | Prompt caching cuts costs by up to 90% | $0.05 per 1M cached input tokens | Cache-hit rate billed separately (see pricing section) |
| Batch discount | 50% | 50% (Batch and Flex) | Not confirmed |
| Open weights | No, hosted only | No, hosted only | Yes, on Hugging Face (checkpoint about 510 GB) |
| Consumer access | Claude.ai Free, Pro, Max, Team, Enterprise (web, iOS, Android) | ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, Edu | DeepSeek website and mobile app |
| Cloud availability | Claude Platform, AWS, Google Cloud, Microsoft Foundry | OpenAI API | DeepSeek API, Hugging Face weights |
| Regional note | Not stated in our sources | No UAE inference residency. Regional processing adds 10% | Not stated in our sources |
| Predecessor | Claude Haiku 4.5 | GPT-5.6 Terra | DeepSeek V4.1 Flash (September 10, 2026) |
Two rows deserve a second look. The long-prompt row shows Anthropic charging 5x more per token once a prompt passes 100,000 tokens, yet Anthropic says about 90% of requests to its previous Haiku fell under that line. And the open-weights row is the only place where DeepSeek stands apart in kind and not just in degree.
Release Timeline: Three Launches in Three Weeks
The order of events matters, because each vendor reacted to a market that moved weekly. DeepSeek set the tone in the late summer. V4 Preview appeared on April 24, 2026, V4-Pro reached general availability on August 13, and a vision experiment, V4-Flash-Vision-Exp, followed on August 21. Then V4.1-Flash landed on September 10 with native multimodal support in the API, and its production pricing took effect at 04:00 UTC the same day.
Four days later, on September 14, DeepSeek began routing requests for deepseek-v4-pro to V4.1-Flash at V4.1-Flash rates, pending the launch of V4.1-Pro. That detail trips up teams that hard-coded the Pro model name. If your bill dropped in mid-September without any change on your side, this routing is the likely reason, and so is the quality shift you may have noticed. Our earlier look at GPT-5.6 against DeepSeek V4 Pro covers the model that was routed away.
Anthropic moved next. Claude Haiku 5.5 arrived earlier this month with a price cut that Anthropic describes as roughly 75% lower on average than Haiku 4.5, and about 90% lower for prompts up to 100,000 tokens. The same announcement halved cache-read pricing on Claude Sonnet 5.5, which Anthropic says makes Sonnet about 20% cheaper for most agentic work, and it added monthly API credits for Max and Team subscribers. Our launch-day pricing coverage of Haiku 5.5 has the first-day reaction.
DeepSeek answered on October 17 with V4.1-Flash-1017, kept the deepseek-flash alias and left the API pricing unchanged. It also opened its Search and Fetch APIs to all users that day. Three days later it released DeepInfer, an open-source inference framework that supports the 1017 model. OpenAI closed the run on October 26 with Terra and published a system card addendum for it the same day.
| Date (2026) | Event | Why it matters for the bill |
|---|---|---|
| April 24 | DeepSeek V4 Preview | Start of the V4 line |
| August 13 | DeepSeek V4-Pro general availability | Premium DeepSeek tier before Flash took over routing |
| September 10 | DeepSeek V4.1-Flash with native multimodal API | New production pricing took effect at 04:00 UTC |
| September 14 | deepseek-v4-pro requests routed to V4.1-Flash | Pro traffic billed at Flash rates until V4.1-Pro ships |
| October 7 | Claude Haiku 5.5 | $0.10 input and $0.50 output up to 100,000 tokens |
| October 17 | DeepSeek V4.1-Flash-1017 | Alias deepseek-flash upgraded, pricing unchanged |
| October 26 | GPT-6.1 Terra | $1 input, $0.05 cached, $5 output |
DeepSeek API Pricing vs Haiku 5.5 vs Terra: Token Costs
Start with the two prices nobody disputes. OpenAI’s page lists GPT-6.1 Terra at $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens on Standard. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold, Haiku 5.5 costs $0.50 input and $2.50 output.
| Price per 1M tokens | Input | Cached input | Output | Source status |
|---|---|---|---|---|
| Claude Haiku 5.5 (up to 100K-token prompts) | $0.10 | Caching saves up to 90% | $0.50 | Anthropic, official |
| Claude Haiku 5.5 (over 100K-token prompts) | $0.50 | Caching saves up to 90% | $2.50 | Anthropic, official |
| GPT-6.1 Terra (Standard) | $1.00 | $0.05 | $5.00 | OpenAI, official |
| GPT-6.1 Terra (Batch or Flex, derived) | $0.50 | $0.025 | $2.50 | OpenAI states Batch and Flex are 50% below Standard |
| DeepSeek V4.1 Flash 1017 | See note below | See note below | See note below | Changelog says pricing unchanged, rate table not confirmed |
The derived Terra row is arithmetic, not a new claim. OpenAI says Batch and Flex prices run 50% below Standard, so $1 becomes $0.50 and $5 becomes $2.50. That puts Terra’s discounted price equal to Haiku 5.5’s price on prompts over 100,000 tokens, which is a useful coincidence for planning. It also means Haiku 5.5 on Standard is still five times cheaper on input than Terra on Batch.
Why the DeepSeek line has a note instead of a number
DeepSeek’s changelog says the API calling method and pricing stayed the same after the October 17 update, and developers reach the new version through deepseek-flash. What the changelog result does not expose is the per-token table. Third-party trackers fill the gap with conflicting figures. One tracker lists $0.14 input and $0.55 output through a provider. Another presents $0.30 and $1.20 as the official price. A third source reports $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, with peak pricing of $0.30 and $1.20, plus cache hits at $0.003 off-peak and $0.006 at peak. Treat all of these as reports and not as a price sheet. Check DeepSeek’s own V4.1 Flash announcement and the live pricing page before you budget.
If the off-peak report holds, DeepSeek would sit slightly above Haiku 5.5 on input and slightly above it on output, and it would double during peak hours. That would reverse the usual assumption that DeepSeek is the cheapest option. It also shows why a time-of-day pricing model matters: a batch job that runs overnight in UTC pays one rate, and a user-facing chat feature in US business hours may pay the other. We include those figures for planning only. For a deeper look at how the older DeepSeek tiers compared with Anthropic’s smaller models, see our Claude Haiku, Gemini Flash and GPT-5.4 Mini price comparison for the previous generation.
Caching, Batch and Long-Prompt Pricing Compared
List price is the headline, but agent workloads rarely pay it. Agents resend the same system prompt, tool definitions and conversation history on every turn, so the cached-input rate decides the real cost. Here the three vendors diverge in how much they tell you.
Prompt caching
OpenAI publishes a clear number: cached input on Terra costs $0.05 per million tokens, 5% of the uncached input rate. Cache writes cost $1.25 per million tokens with 30-minute retention and $2.50 with 12-hour retention, before long-context adjustments. Anthropic says prompt caching on Haiku 5.5 can cut costs by up to 90% but did not put a per-token cache-read figure on the page our research retrieved. It did halve the cache-read price on Sonnet 5.5, which signals how seriously it treats agentic workloads. DeepSeek, per the third-party report, bills cache hits at a fraction of a cent.
The practical takeaway: if more than half of your input tokens are repeated context, run your own cost model with the vendors’ live cache tables. A model with a higher list price and a very low cache rate can beat a cheaper-looking rival once caching is on.
Batch and flex tiers
Both Western vendors give 50% off for asynchronous work. OpenAI applies it to Batch and Flex. Anthropic applies it to batch processing. If you run nightly classification, bulk summarization or evaluation sweeps, the discount stacks with the low base price and brings the cost of a million short requests down to a few dollars. Teams that run on AWS can also compare how a cloud provider layers its own tiers on top. Our guide to Amazon Bedrock pricing tiers shows how Flex, Priority and Batch differ on that platform.
Long prompts and regional processing
Anthropic’s split at 100,000 tokens makes long-document work five times more expensive per token on both input and output, though still half of Terra’s Standard rate. OpenAI adds a 10% premium for regional processing where available and says Terra does not support inference residency in the UAE. If your compliance team needs in-region processing, check that line before you pick a model, because it can remove Terra from the shortlist in a single sentence.
Four Worked Cost Scenarios at Real Volumes
Prices only help when you apply them to a workload. The scenarios below use the verified Standard rates from the two Western vendors and plain arithmetic. They assume no caching unless noted, and they use Haiku 5.5’s rate for prompts up to 100,000 tokens except in the long-document case.
| Scenario | Volume | Haiku 5.5 | GPT-6.1 Terra | Gap |
|---|---|---|---|---|
| A. Support bot | 1M requests, 1,500 in and 300 out tokens each | $300 ($150 in, $150 out) | $3,000 ($1,500 in, $1,500 out) | 10x |
| B. Document summarizer | 200,000 docs, 6,000 in and 400 out each | $160 ($80 on Batch) | $1,600 ($800 on Batch) | 10x |
| C. Coding agent loop | 50,000 runs, 40,000 in and 2,000 out each | $250 | $2,500 ($980 with 80% cache hits) | 10x, or about 4x with caching |
| D. Long-document review | 1,000 requests, prompts over 100,000 tokens each | Not listed | Not listed | Not listed |
Scenario A is the classic chatbot. A million requests at 1,500 input and 300 output tokens each means 1,500 million input tokens and 300 million output tokens. Haiku 5.5 charges $150 for the input and $150 for the output. Terra charges $1,500 for each. A team running that volume saves $2,700 a month by choosing Haiku, and the saving scales linearly.
Scenario B shows the batch effect. Summarizing 200,000 documents of 6,000 tokens each is almost entirely input. Haiku costs $160 on Standard and $80 on Batch. Terra costs $1,600 and $800. The ratio stays at 10x because both vendors apply the same 50% discount.
Scenario C is where Terra fights back. A coding agent that resends 40,000 tokens of context across 50,000 runs is mostly repeated input. At 80% cache hits, the cached share bills at $0.05 per million tokens and the rest at $1 per million, plus output at $5 per million, for $980 in total. Haiku 5.5 without any caching discount costs $250. The gap narrows from 10x to about 4x, and Haiku’s own caching would shrink it further, though Anthropic has not published the per-token cache figure in the material we reviewed.
Scenario D covers legal, research and compliance review, where prompts pass 100,000 tokens. For those long prompts, Haiku 5.5 bills $0.50 per million input tokens and $2.50 per million output tokens. Terra bills $1 and $5 per million on its standard list price. The gap falls to 2x, so long prompts are where Terra’s premium is easiest to justify.
What about DeepSeek? One report puts off-peak pricing at $0.15 input and $0.60 output per million tokens, with peak pricing at $0.30 input and $1.20 output. Those figures are unconfirmed, and we flag them as a planning range and not a quote, so we have not priced Scenario A on them.
Benchmarks: What Each Vendor Actually Published
Here is the honest state of the evidence on October 27. No independent head-to-head test of all three models has landed yet, and the vendors picked different benchmarks. That is why this section maps coverage and not a single leaderboard. A table of invented scores would look tidy and mislead you, so the table below shows who cites what, and the text reports every number we could trace to a source.
| Benchmark | What it measures (per vendor descriptions) | Cited by |
|---|---|---|
| Terminal-Bench 4.0 | Complex, multi-step professional tasks in a command-line interface | Anthropic (Haiku 5.5 page), DeepSeek (run with the Claude Code framework) |
| FrontierCode 1.1 | Agentic coding | Anthropic |
| OSWorld 2.1 | Operating a real computer to finish long, multi-step tasks | Anthropic |
| MMMU-Pro | Visual reasoning | OpenAI (Terra page) |
| Humanity’s Last Exam (text subset) | Hard academic questions, no tools | DeepSeek |
| DeepSWE v1.1 | Software engineering tasks, run with the Mini-SWE-V2 framework | DeepSeek |
| OpenAI production benchmarks | Internal tests in the Terra system card addendum | OpenAI |
Source 1: OpenAI on GPT-6.1 Terra
OpenAI says Terra beats GPT-5.6 Terra on MMMU-Pro at every reasoning setting and nearly matches the scores of GPT-6 Astra and GPT-6.1 Sol at much lower cost. The system card addendum adds that Terra performs on par with GPT-5.6 Terra on the production benchmarks it cites. Read that sentence twice. It describes a cost-down move, not a capability jump over the previous Terra. Digital Trends framed the launch as Astra-like results at a fraction of the cost, which matches OpenAI’s positioning. The same addendum reportedly rates Terra Critical for cybersecurity and High for biological and chemical capability under OpenAI’s Preparedness Framework, a single-source detail that matters for enterprise review boards.
Source 2: Anthropic on Claude Haiku 5.5
Anthropic’s page shows charts for Terminal-Bench 4.0, FrontierCode 1.1 and OSWorld 2.1, with accuracy plotted against cost. The text we retrieved does not give the numeric scores, so we do not repeat any. What the page does state in words is that Haiku 5.5 is a significant step up over Haiku 4.5 across coding, tool use, computer use and agents. For a baseline, Anthropic’s Haiku 4.5 page reports 73.3% on SWE-bench Verified, a figure that now describes the predecessor. Anthropic also notes that Haiku 5.5 is its fastest model at standard speed but runs less quickly than its Opus models in Fast Mode.
Source 3: DeepSeek on its own 1017 revision
DeepSeek reports gains over the September V4.1 Flash. According to its release notes, Humanity’s Last Exam without tools rose from 36.8 to 42.6. Terminal-Bench 4.0 rose from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 from 74.2 to 77.0. Those are vendor-reported numbers, and DeepSeek’s footnote says the HLE run used only the pure-text subset, Terminal-Bench 4.0 used the Claude Code framework and DeepSWE used Mini-SWE-V2. Quartz covered the update under a headline about better scores. The gains are large for a point release, which is exactly why independent replication matters.
Source 4: independent safety testing
One independent group tested the DeepSeek model and reported stronger safety performance than earlier versions. It also found failures in some shutdown-resistance and long-task evaluations, according to coverage of the tests. That mixed result belongs in any procurement file, especially for agents that run unattended.
The only benchmark that appears in two vendors’ material is Terminal-Bench 4.0, and even there the harnesses differ. For frontier-tier context on the same family names, our LLM leaderboard covering Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon lists the bigger models these small ones trail.
Coding and Agent Work: Subagents, Harnesses and Tool Use
Small models earn their keep in coding workflows as subagents. A larger model plans the work, then hands narrow tasks to a cheap model that edits files, runs commands or searches a codebase. Anthropic describes Haiku 5.5 in exactly those terms. A smarter model such as Opus can plan and delegate, and Haiku can run many agents in parallel. The page also says Haiku 5.5 handles focused coding and multi-step tool use, like direct edits and small changes applied across many files. It is available in Claude Code, so the delegation pattern needs no extra glue.
Terra targets a similar slot inside OpenAI’s products. It is rolling out in ChatGPT Work and Codex, and developers can call it through the API as gpt-6.1-terra. OpenAI’s release notes add that overall plan usage limits stay unchanged, and that some Plus and Pro users may see Terra at Medium instead of GPT-6.1 Sol at Light in the default Work power presets. For a Codex user, that means the cheaper model may quietly take over tasks the heavier one handled before. Check your power preset if output quality shifts after the rollout.
DeepSeek leans into the agent harness ecosystem. The release notes say the model is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, among others. DeepSeek also gave server-side Search and Fetch tool calls through DeepSeek Harness, Claude Code and Codex for free until 16:00 UTC on October 31, 2026. After that, each endpoint costs $0.75 per 1,000 calls, a price that PPC Land reported alongside news about bot conduct grading. If your agents browse heavily, add that line item to your model, because tool calls can cost more than the tokens.
Which one writes better code? The evidence does not settle it. DeepSeek’s Terminal-Bench 4.0 run used the Claude Code framework, Anthropic reports on FrontierCode 1.1, and OpenAI cites its own production benchmarks. Until a neutral lab runs all three on one harness, pick by workflow fit and price. Our MiniMax, GPT-6 Luna and DeepSeek V4.1 Flash comparison tests another price-tier trio if you want a wider field, and the AI and machine learning section tracks new matchups as they land.
Computer Use, Vision and Multimodal Tasks
All three vendors now pitch their small model as capable of operating software, not only writing text. OpenAI lists computer use and image analysis in Terra’s positioning, and its API page suggests matching product photos to catalog entries as a use case. Anthropic calls Haiku 5.5 a strong computer use agent for repetitive tasks like form filling, data entry and moving information between apps, and says it is cost efficient at scale. DeepSeek added native multimodal support to the API with V4.1 Flash on September 10, after the experimental vision model in August.
For computer-use agents the cost argument sharpens. A browser agent takes many steps, and each step may send a screenshot plus history. That pattern looks like Scenario C above, with many input tokens per run and few output tokens. At Haiku 5.5’s $0.10 input rate, a long agent session costs cents. At Terra’s $1 rate the same session costs ten times more unless the cache hit rate is high. xenospectrum framed Terra’s launch around cheaper computer use, so OpenAI clearly sees the same economics.
OSWorld 2.1, which Anthropic cites, measures how well agents operate a real computer on long tasks, and Anthropic plots accuracy against cost on it. Terra’s MMMU-Pro claim tests a different skill, visual reasoning on static images. Do not read one as a proxy for the other. If your workload is screenshot-driven automation, run a 50-task pilot on both models with your own screens, because UI layouts, fonts and login flows vary far more than any public benchmark assumes.
Open Weights, Self-Hosting and Data Control
DeepSeek is the outlier here. It publishes V4.1-Flash-1017 on Hugging Face as an open-weight release, and the reported checkpoint size is about 510 GB. That size rules out a single consumer GPU, but it fits a multi-GPU server. DeepSeek also released DeepInfer, an open-source inference framework that supports the model. For teams in regulated industries, self-hosting is the only way to keep prompts entirely inside your own network.
Self-hosting changes the cost question from price per token to price per GPU hour. A 510 GB checkpoint needs serious hardware, so the break-even point sits at sustained, high utilization. Teams that call the model a few thousand times a day will spend less on a hosted API. Teams that process hundreds of millions of tokens daily, with steady load, can come out ahead on their own cluster. Our guide to the best local LLMs covers the hardware side of that decision.
The hosted options differ on control, too. OpenAI says Terra has no UAE inference residency and charges a 10% premium for regional processing where it is available. Anthropic offers Haiku 5.5 through AWS, Google Cloud and Microsoft Foundry, which lets enterprises use existing cloud contracts, private networking and regional controls. DeepSeek’s hosted API is the option that raises the most data-governance questions for US and EU enterprises, and the independent safety findings add to that review burden.
A last point for compliance leads. Open weights mean you can audit and pin a version. DeepSeek’s hosted alias deepseek-flash moved from the September build to the 1017 build without a code change on your side. That convenience is also a risk. If you need reproducible outputs, pin the open-weight checkpoint or log the model version on every request.
Real-World Examples: Where These Models Already Run
These are documented deployments and distribution points from the first three weeks, not hypothetical pilots.
- Kiro, the agentic IDE. The Kiro team announced that Claude Haiku 5.5 is available in Kiro, with reports that it reached the IDE, CLI, web and Crew surfaces. It shows a small model landing directly in a developer tool.
- Claude Code subagents. Haiku 5.5 is available in Claude Code. Anthropic positions it as the fast worker that a larger model delegates to, so one session can fan out many parallel agents at $0.10 per million input tokens.
- Codex and ChatGPT Work. Terra rolled into Codex and ChatGPT Work on October 26 and 27 for Plus, Pro, Business, Enterprise and Edu plans. ChatGPT Work users may meet it without choosing it, through the default power presets.
- DeepSeek inside rival harnesses. DeepSeek made Search and Fetch tool calls free through DeepSeek Harness, Claude Code and Codex until October 31. A DeepSeek model can run inside the same tools that front Anthropic and OpenAI models.
- Community quantization on Hugging Face. Within days of the 1017 release, an independent contributor published a quantized GGUF variant. Open weights turn a launch into an ecosystem almost at once.
- Subscriber credits for Haiku 5.5. Anthropic gave Max 5x subscribers $100 in monthly API credits, Max 20x subscribers $200 and Team plans up to $500 pooled across users. A solo developer on Max can run substantial Haiku workloads at no extra marginal cost.
- Multi-cloud enterprise rollout. Haiku 5.5 shipped on AWS, Google Cloud and Microsoft Foundry on day one, so enterprise buyers can route it through existing procurement.
Hands-on developers who want to call the DeepSeek side directly can follow our DeepSeek V4.1 Flash API setup guide, which walks through keys, endpoints and a first request.
Use-Case Recommendations: Which Model for Which Job
The right pick depends on the shape of the workload. Here are six patterns and the model that fits each on the evidence available today.
| Use case | Best fit | Reason |
|---|---|---|
| High-volume support chat | Claude Haiku 5.5 | Lowest verified price. Anthropic lists live customer support as a target use |
| Bulk summarization and classification | Claude Haiku 5.5 on Batch | Roughly $80 for 200,000 documents in Scenario B |
| Coding subagents | Claude Haiku 5.5 or Terra | Haiku inside Claude Code, Terra inside Codex. Match your primary tool |
| Long-document review above 100K tokens | Terra if quality needs it, Haiku if cost does | Gap shrinks to 2x in Scenario D |
| Screenshot and image-heavy automation | Pilot Terra and Haiku | Terra claims strong MMMU-Pro scores. Haiku targets computer use. Test both |
| Self-hosted or air-gapped deployments | DeepSeek V4.1 Flash 1017 | Only open-weight option, about 510 GB checkpoint |
Three extra notes. First, if your team lives in the OpenAI ecosystem and already pays for Plus, Pro or Business, Terra sits inside your existing plan limits in ChatGPT Work and Codex, and the API price is reasonable for moderate volume. Second, if you need a model with more reasoning headroom, step up to the mid tier. Our Sonnet 5.5, GPT-6 Luna and Gemini 3.8 comparison lays out that tier. Third, for any workload that sends personal data, check the residency and governance rows in the specs table before the price rows.
Pros and Cons of Each Model
Claude Haiku 5.5
- Pro: Lowest verified list price at $0.10 input and $0.50 output for prompts up to 100,000 tokens.
- Pro: Available on the Claude Platform, AWS, Google Cloud and Microsoft Foundry, plus Claude.ai and Claude Code.
- Pro: Max and Team subscribers get monthly API credits.
- Con: Price jumps 5x per token above 100,000 tokens.
- Con: Closed weights. You cannot self-host or pin a local copy.
- Con: Anthropic publishes charts for its benchmarks, and the page we reviewed gave no numeric scores in text.
GPT-6.1 Terra
- Pro: Very low cached-input rate at $0.05 per million tokens, which suits agents.
- Pro: Strong visual reasoning claims on MMMU-Pro, close to Astra and Sol at lower cost per OpenAI.
- Pro: Included in ChatGPT Work and Codex for paid plans.
- Con: Standard price is 10x Haiku 5.5 on both input and output.
- Con: On par with GPT-5.6 Terra on OpenAI’s production benchmarks, so it is a cost move more than a capability jump.
- Con: No UAE inference residency, and a 10% premium for regional processing.
DeepSeek V4.1 Flash 1017
- Pro: Open weights on Hugging Face, with community quantizations and an open-source inference framework.
- Pro: Large vendor-reported gains over the September build on agentic coding tests.
- Pro: Stable alias keeps integrations simple and pricing was unchanged by the update.
- Con: Confirmed per-token rate sheet is hard to pin down, and reported peak pricing is double off-peak.
- Con: Independent testing found failures in some shutdown-resistance and long-task evaluations.
- Con: Alias upgrades change behavior without a code change, which complicates reproducibility.
Migration Guide: Switching Between the Three Models
Most teams will not rip out an existing model. They will add a second model for the cheap, high-volume traffic and keep the first for hard cases. This five-step plan keeps the switch reversible.
- Log the baseline. Record input tokens, output tokens, latency and a quality score for 200 to 500 real requests from your current model. Without this baseline, no benchmark chart will tell you whether the switch helped.
- Put the model name behind a config value. Never hard-code
deepseek-v4-proor any other name. The September 14 routing change showed how a name can point to different hardware without your consent. - Shadow-test the new model. Send the same prompts to the old and new models, store both answers, and compare with a rubric or a judge model. Run at least a few hundred cases per task type.
- Move low-risk traffic first. Start with classification, tagging or draft summaries, where a wrong answer is cheap. Keep the larger model for customer-facing replies until your numbers support the move.
- Set a rollback trigger. Define a quality or error-rate threshold in advance. If the new model crosses it, flip the config back.
The snippet below shows a thin router that can target all three providers through one function. The model identifiers come from the vendors’ pages: claude-haiku-5-5, gpt-6.1-terra and deepseek-flash. DeepSeek’s endpoint accepts OpenAI-style requests, so the OpenAI client works with a different base URL. Check the current docs for each SDK version before you ship.
import os
from anthropic import Anthropic
from openai import OpenAI
MODELS = {
"haiku": {"vendor": "anthropic", "model": "claude-haiku-5-5"},
"terra": {"vendor": "openai", "model": "gpt-6.1-terra"},
"deepseek": {"vendor": "deepseek", "model": "deepseek-flash"},
}
anthropic_client = Anthropic()
openai_client = OpenAI()
deepseek_client = OpenAI(
api_key=os.environ["DEEPSEEK_API_KEY"],
base_url="https://api.deepseek.com",
)
def ask(choice: str, prompt: str, max_tokens: int = 500) -> str:
cfg = MODELS[choice]
if cfg["vendor"] == "anthropic":
r = anthropic_client.messages.create(
model=cfg["model"],
max_tokens=max_tokens,
messages=[{"role": "user", "content": prompt}],
)
return r.content[0].text
client = openai_client if cfg["vendor"] == "openai" else deepseek_client
r = client.chat.completions.create(
model=cfg["model"],
max_tokens=max_tokens,
messages=[{"role": "user", "content": prompt}],
)
return r.choices[0].message.contentA second snippet turns the verified prices into a cost estimator, so your finance team can check Scenario A-style math against live traffic. Update the table whenever a vendor changes its page.
PRICES = { # USD per 1M tokens: (input, output)
"haiku_short": (0.10, 0.50), # prompts up to 100K tokens
"haiku_long": (0.50, 2.50), # prompts over 100K tokens
"terra": (1.00, 5.00), # Standard
"terra_batch": (0.50, 2.50), # 50% below Standard
}
def monthly_cost(key, requests, in_tokens, out_tokens):
pin, pout = PRICES[key]
return requests * (in_tokens * pin + out_tokens * pout) / 1_000_000
print(monthly_cost("haiku_short", 1_000_000, 1500, 300)) # 300.0
print(monthly_cost("terra", 1_000_000, 1500, 300)) # 3000.0When you move from the OpenAI client to a tool-calling workflow, schemas and tool-choice behavior differ between vendors. Our function calling tutorial across four LLMs covers those differences step by step. Last, if you used deepseek-v4-pro before September 14, confirm which model actually answered your requests. The routing note in the timeline above explains why the answer may be Flash.
Verdict: Haiku 5.5 Wins on Price, Terra on Fit, DeepSeek on Control
Pick Claude Haiku 5.5 as the default for high-volume work. It has the lowest confirmed price at $0.10 input and $0.50 output, it ships on three major clouds, and Anthropic aims it squarely at subagents, support and computer use. In Scenario A it cuts a $3,000 monthly bill to $300. Only the 100,000-token threshold and the closed weights hold it back.
Pick GPT-6.1 Terra when your workflow already lives in OpenAI’s products, when image analysis and visual reasoning matter, or when cache-heavy agent loops dominate your traffic. The $0.05 cached rate narrows the gap with Haiku from 10x to roughly 4x in Scenario C. But OpenAI’s own addendum puts Terra on par with GPT-5.6 Terra, so treat it as a cheaper way to get the old quality and not as a new ceiling.
Pick DeepSeek V4.1 Flash 1017 when you need open weights, a self-hosted path or the lowest reported rates. It is the only one of the three you can run on your own hardware, and its vendor-reported benchmark gains are large. The cost of that freedom is a murky rate sheet, a time-of-day pricing report, an alias that changes under you and an independent safety result that mixes good news with failures. Pin the version and verify price before launch.
The cleanest strategy for most teams is a router. Send routine traffic to Haiku 5.5, send cache-heavy or image-heavy jobs to Terra, and keep DeepSeek ready for workloads that need local control. Re-run your own 500-request test every time a vendor ships a new checkpoint, which at the current pace means about once a month.
Frequently Asked Questions
What is the cheapest model of the three?
On confirmed list prices, Claude Haiku 5.5 is cheapest at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. GPT-6.1 Terra costs $1 and $5. DeepSeek’s per-token table is not confirmed in our sources, though third-party reports put its off-peak rates in the same range as Haiku.
When was GPT-6.1 Terra released?
OpenAI introduced GPT-6.1 Terra on October 26, 2026. It rolled out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu plans, and developers can call it through the API as gpt-6.1-terra.
Did DeepSeek change its API price with the 1017 update?
DeepSeek’s changelog says the API calling method and pricing stayed the same. Use the alias deepseek-flash to reach the new version. Search and Fetch calls will cost $0.75 per 1,000 calls per endpoint once free access through DeepSeek Harness, Claude Code and Codex ends on October 31, 2026, at 16:00 UTC.
Can I run DeepSeek V4.1 Flash 1017 myself?
Yes. DeepSeek published the weights on Hugging Face. The checkpoint is about 510 GB, so it needs a multi-GPU server, and DeepSeek’s open-source DeepInfer framework supports the model. An independent contributor has already posted a quantized GGUF build.
Is Claude Haiku 5.5 better than Haiku 4.5?
Anthropic says Haiku 5.5 is a significant step up over Haiku 4.5 across coding, tool use, computer use and agents, and that it costs around 75% less to run on average. For prompts up to 100,000 tokens, the saving is about 90%.
Which model should I use for coding agents?
No neutral test has run all three on one harness. Haiku 5.5 is available in Claude Code and fits the subagent role. Terra runs in Codex. DeepSeek reports large gains on Terminal-Bench 4.0 using the Claude Code framework. Pilot on your own repositories and pick by measured pass rate per dollar.
Does GPT-6.1 Terra support data residency in every region?
No. OpenAI lists no UAE inference residency for Terra, and regional processing adds a 10% premium where it is available. Check the current residency list for your region before committing.
Will these prices change soon?
Probably. All three vendors moved prices or shipped new checkpoints within a 20-day window. Treat every price in this article as valid on October 27, 2026, and confirm on the vendor pages before you sign a budget.
