Does the $200 Codex plan suck now?

Core Complaint: the $200 Plan Got Much Worse Value

The video argues that OpenAI’s and Anthropic’s long-running “$200 plan as near-unlimited” bargain deteriorated sharply after major changes to usage metering and after new models/tier changes.

1) The core complaint: $200 delivers far less inference value now

• The creator claims that a few weeks earlier, the $200 Codex/Claude-style subscription could translate into extremely high inference usage (citing figures like ~$12K/month for Codex and ~$8K/month on Claude).
• After OpenAI’s “Dev Day” change (attributed to “Tibo”), the way usage is calculated for the $200 plan effectively halves the value of the allowance.
• He reports his recalculated/observed result:
  • About $570 worth of usage over a week (roughly $2K/month), a major drop from earlier expectations.
• He questions whether there’s any reason left to use Codex vs. Claude subscriptions, or whether the change is effectively pushing users toward the $500 tier.

Why This Happened: Subsidy Economics and a “Subsidy Arms Race”

2) Changing industry economics

The creator frames the situation as an industry-wide competition:

• Companies didn’t want $200 users to feel they were getting “unlimited” value once newer models arrived.
• So OpenAI and Anthropic increased subsidization over time to preserve perceived value—eventually creating extreme pricing dynamics.

2.1) Misleading “value” comparisons

He argues that both the advertised “value” numbers and many API comparison metrics are misleading because:

• Subscriptions have very high margins relative to true compute/energy costs.
• The “outlandish gap” is driven more by pricing strategy and enterprise/API charge-offs than by real cost.

He also suggests that increased competition could eventually compress those gaps (through lower enterprise/API prices), but for now consumers experience worse effective value.

Codex vs. Claude Plans: Restrictions and Model Handling

3) Claude subscriptions may be operationally worse (despite appearing cheaper)

The creator claims Claude’s subscription is often worse in day-to-day use, even if it looks cheaper on paper:

• Claude sub can only be used in “Claude code” (with narrow exceptions via specific integrations like T3).
  • He claims you can’t easily route the subscription through other tooling/services (e.g., Pi/OpenRouter/“Hermes”).
  • He says attempts to “trick prompts” won’t work.
• OpenAI Codex sub is more permissive:
  • Usable for many non-production/engineering workflows (e.g., CI).
  • The main restriction is avoiding use to serve end-user traffic.
• He also claims Claude restricts or doesn’t fully subsidize usage for its biggest models, contrasting:
  • “Opus-like” vs “Fable-like” expensive model groupings.

Model Pricing Analysis: Margins Have Dropped

4) Pricing/margin argument: less subsidy means tighter allowance

A large portion of the video focuses on pricing and margin shifts.

• Anthropic (Fable 5 / “Mythos”)

  • The creator says the pricing changed relative to earlier expectations.
  • He claims the same hosting/model size now implies lower margins than prior assumptions (e.g., ~95%).
  • Net effect: Anthropic can still be profitable, but they must tighten subscription subsidization because they’re “printing less money per compute.”

• OpenAI (6.1 Soul)

  • He argues OpenAI’s 6.1 Soul pricing suggests worse margins than users expected.
  • Implication: these newer models are likely being subsidized less aggressively than before.

Personal “Real Number” Results: ~10x Drop in Effective Usage

5) Measured usage outcomes

He reports his own measured usage:

• Previously: around ~$24K/month in a “maxed out” scenario during subsidy-heavy windows.
• Now: closer to ~$2.5K/month on the $200 plan—about a 10x drop.

5.1) Astra vs. 6.1 Soul economics

He supports this by contrasting:

• Astra: ~766K output tokens costing about $534 (API-priced).
• 6.1 Soul: similar output token volume costing about $38.

His conclusion: the per-token economics are far better for Soul, making it the key to restoring “worth it” subscription behavior.

Value Conclusions (and Criticism of Trust/Optics)

6) Value vs. trust/branding

He distinguishes between two ideas:

• Value

  • Calls 6.1 Soul the best value model by far.
  • Says Opus/Claude tiers still have strong engineering usefulness, especially for shipping code.

• Trust / branding / communication

  • He believes OpenAI’s rollout order and tier messaging harmed goodwill.
  • Even though Soul improves efficiency and can restore “unlimited-feeling” behavior if you use the right model, the experience is still worse because the changes weren’t clearly communicated early enough.

The $500 Tier and “Ultrafast” Issues

7) Optics: the $500 rollout is described as a major mistake

The creator argues the $500 tier rollout was the biggest optically damaging error:

• Branding it as a premium upgrade while the perceived “20x → 10x” experience is worse for many users.
• He argues subscription tiers are mostly marketing, not the true profit engine—so optics matter.

7.1) “Ultrafast” is portrayed as especially expensive

He claims:

• Ultrafast is extremely expensive and designed to burn limits quickly.
• He provides examples suggesting that reviewing only a couple PRs on Ultra fast can cost hundreds (e.g., “$600”).
• Even the release timing is alleged to be problematic:
  • 61 Soul efficiency gains were undermined by simultaneous changes that reduce effective allowances.
• He alleges users upgrading for Ultrafast can hit account limits within ~an hour or two.

Final Answer: Does the $200 Plan “Suck” Now?

8) Depends on why you bought it—and which model you rely on

• If users bought the $200 plan mainly for Astra, the creator says: yes, it sucks now.
• If users can rely on 6.1 Soul, he says the subscription can still deliver strong output-per-dollar and isn’t a “rug pull.”
• He says Claude remains the better option for serious regular engineering work, but also argues OpenAI models may be more thorough for certain analysis workflows when Codex is used via Claude agents.

---

Presenters / Contributors

• Main presenter/creator: The video narrator/author (speaking throughout; not named in the provided subtitles)
• Sponsor/ads mentioned: Browserbase; Sywand / General Translation (referenced as “General Translation” with swyd.link/gt)
• Other referenced individuals/companies (not presenters):
  • Tibo (referenced regarding the announcement)
  • OpenAI, Anthropic
  • Jane Street (mentioned regarding chips)
  • Model/tool ecosystem: Codex, Claude, Sonnet, Opus, Astra, Fable, Soul, 6.1, Ultrafast, and related tooling.